Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Nature Medicine (2026)Cite this article Growing evidence points to a close neurophysiological link between brain and body. Recent rodent studies have shown that the dopaminergic mesolimbic pathway, which underlies expectations of positive outcomes, also modulates immune function. However, it remains unknown whether a similar brain-immune link exists in humans and whether it involves conscious positive expectations. In a preregistered, double-blind randomized controlled trial, we used fMRI neurofeedback (NF) to train healthy participants to increase reward mesolimbic activity through self-chosen mental strategies, followed by an immune challenge with the hepatitis B virus (HBV) vaccine and assessments of HBV antibody (HBVab) levels. Eighty-five participants were randomized to (1) reward mesolimbic upregulation (n = 34), (2) non-mesolimbic control upregulation (n = 34) or (3) no-NF control (n = 17). Prespecified primary outcomes were (1) differences in reward mesolimbic activation between NF groups, (2) correlation between reward mesolimbic upregulation and post-vaccination HBVab changes across both NF groups and (3) group differences in post-vaccination HBVab changes. Both NF groups showed significant increases in reward mesolimbic activation. Notably, greater ventral tegmental area (VTA) upregulation—but not nucleus accumbens or control region activation—was associated with larger post-vaccination increases in HBVab levels (r = 0.31, P = 0.018). Sustained VTA upregulation was further linked to mental strategies involving positive expectations. Post-vaccination antibody levels did not differ between groups, and no adverse effects occurred. Together, these findings suggest that consciously generated positive expectations can engage reward circuitry to influence immune function, a process that may be leveraged for non-invasive immune modulation. This is a preview of subscription content, access via your institution Access Nature and 54 other Nature Portfolio journals Get Nature+, our best-value online-access subscription cancel any time Subscribe to this journal Receive 12 print issues and online access $259.00 per year only $21.58 per issue Buy this article Prices may be subject to local taxes which are calculated during checkout Requests for de-identified data can be directed at the corresponding author (T.H., talma@tlvmc.gov.il). All requests for data sharing will be reviewed by the Tel Aviv Sourasky Medical Center IRB committee to verify whether the request is subject to any intellectual property or confidentiality obligations. Requests will be reviewed on the basis of scientific merit, ethical review, available resources and regulatory requirements and will be responded within 90 days. After approval of a proposal, anonymized individual-level data will be made available for reuse in accordance with the signed consent IRB form. A signed data access agreement with the collaborator is required before accessing shared data. Analysis codes are available via Github: Wager, T. D. & Atlas, L. Y. The neuroscience of placebo effects: connecting context, learning and health. Google Scholar Gershman, S. J. & Uchida, N. Believing in dopamine. Google Scholar Ben-Shaanan, T. L. et al. Activation of the reward system boosts innate and adaptive immunity. Google Scholar Ben-Shaanan, T. L. et al. Modulation of anti-tumor immunity by the brain's reward system. Google Scholar Kayama, T., Ikegaya, Y. & Sasaki, T. Phasic firing of dopaminergic neurons in the ventral tegmental area triggers peripheral immune responses. Costi, S. et al. Peripheral immune cell reactivity and neural response to reward in patients with depression and anhedonia. et al. Goal-striving tendencies moderate the relationship between reward-related brain function and peripheral inflammation. Brain Behav. Treadway, M. T. et al. Association between interleukin-6 and striatal prediction-error signals following acute stress in healthy female participants. Sitaram, R. et al. Closed-loop brain training: the science of neurofeedback. Lubianiker, N. et al. Process-based framework for precise neuromodulation. Lubianiker, N., Paret, C., Dayan, P. & Hendler, T. Neurofeedback through the lens of reinforcement learning. Trends Neurosci. Klöbl, M. et al. Reinforcement and punishment shape the learning dynamics in fMRI neurofeedback. Oblak, E. F., Lewis-Peacock, J. & Sulzer, J. S. Self-regulation strategy, feedback timing and hemodynamic properties modulate learning in a simulated fMRI neurofeedback environment. PLoS Comput. Zoefel, B., Huster, R. J. & Herrmann, C. S. Neurofeedback training of the upper alpha frequency band in EEG improves cognitive performance. Autenrieth, M., Kober, S. E., Neuper, C. & Wood, G. How much do strategy reports tell about the outcomes of neurofeedback training? A study on the voluntary up-regulation of the sensorimotor rhythm. Qiu, S. et al. Significant transcriptome and cytokine changes in hepatitis B vaccine non-responders revealed by genome-wide comparative analysis. Kirschner, M. et al. Ventral striatal hypoactivation is associated with apathy but not diminished expression in patients with schizophrenia. J. Psychiatry Neurosci. Ohmann, H. A., Kuper, N. & Wacker, J. Examining the reliability and validity of two versions of the Effort-Expenditure for Rewards Task (EEfRT). Gonen, T. et al. Human mesostriatal response tracks motivational tendencies under naturalistic goal conflict. Munshi, S. & Rosenkranz, J. Effects of peripheral immune challenge on in vivo firing of basolateral amygdala neurons in adult male rats. Zhu, X. et al. Somatosensory cortex and central amygdala regulate neuropathic pain-mediated peripheral immune response via vagal projections to the spleen. Boukezzi, S. et al. Exaggerated amygdala response to threat and association with immune hyperactivity in depression. Brain, Behav., Immun. Yin, L. et al. Inflammation and decreased functional connectivity in a widely-distributed network in depression: Centralized effects in the ventral medial prefrontal cortex. Brain Behav. Hui, M. & Beier, K. T. Defining the interconnectivity of the medial prefrontal cortex and ventral midbrain. Jennings, J. H. et al. Distinct extended amygdala circuits for divergent motivational states. Glimcher, P. W. Understanding dopamine and reinforcement learning: the dopamine reward prediction error hypothesis. Natl Acad. Lerner, T. N., Holloway, A. L. & Seiler, J. L. Dopamine, updated: reward prediction error and beyond. Niv, Y. Cost, benefit, tonic, phasic. Beeler, J. Tonic dopamine modulates exploitation of reward learning. Niv, Y., Daw, N. D., Joel, D. & Dayan, P. Tonic dopamine: opportunity costs and the control of response vigor. Hahn, A. et al. Functional dynamics of dopamine synthesis during monetary reward and punishment processing. Blood Flow. Schmidt, *Clemens et al. Multimodal assessment of dopamine synthesis and bold- signaling in reward and punishment processing - a hybrid fpet/fmri study. Article PubMed Central Fields, H. L. & Margolis, E. B. Understanding opioid reward. Trends Neurosci. Laurent, V., Morse, A. K. & Balleine, B. W. The role of opioid processes in reward and decision-making. Bossong, M. G., Wilson, R., Appiah-Kusi, E., McGuire, P. & Bhattacharyya, S. Human striatal response to reward anticipation linked to hippocampal glutamate levels. Benedetti, F. & Amanzio, M. Mechanisms of the placebo response. Zubieta, J.-K. & Stohler, C. S. Neurobiological mechanisms of placebo responses. MacInnes, J. J., Dickerson, K. C., Chen, N. & Adcock, R. A. Cognitive neurostimulation: learning to volitionally sustain ventral tegmental area activation. Totah, N. K. B., Kim, Y. & Moghaddam, B. Distinct prestimulus and poststimulus activation of VTA neurons correlates with stimulus detection. Howe, M. W., Tierney, P. L., Sandberg, S. G., Phillips, P. E. M. & Graybiel, A. M. Prolonged dopamine signalling in striatum signals proximity and value of distant rewards. & Shohamy, D. Intrinsic connectivity between the hippocampus, nucleus accumbens, and ventral tegmental area in humans. Cauda, F. et al. Functional connectivity and coactivation of the nucleus accumbens: a combined functional connectivity and structure-based meta-analysis. Hammes, J. et al. Dopamine metabolism of the nucleus accumbens and fronto-striatal connectivity modulate impulse control. Zunhammer, M., Gerardi, M. & Bingel, U. The effect of dopamine on conditioned placebo analgesia in healthy individuals: a double-blind randomized trial. Kunkel, A. et al. Dopamine has no direct causal role in the formation of treatment expectations and placebo analgesia in humans. PLOS Biol. Lange, T., Dimitrov, S., Bollinger, T., Diekelmann, S. & Born, J. Sleep after vaccination boosts immunological memory. Yu, X. et al. GABA and glutamate neurons in the VTA regulate sleep and wakefulness. Schiller, M., Ben-Shaanan, T. L. & Rolls, A. Neuronal regulation of immunity: why, how and where? Castro, D. C. & Bruchas, M. R. A motivational and neuropeptidergic hub: anatomical and functional diversity within the nucleus accumbens shell. van Dongen, Y. C. et al. Anatomical evidence for direct connections between the shell and core subregions of the rat nucleus accumbens. Cloninger, C. R., Przybeck, T. R. & Svrakic, D. M. The tridimensional personality questionnaire: U.S. normative data. Costa, P. T. & McCrae, R. R. Normal personality assessment in clinical practice: the NEO Personality Inventory. Torrubia, R., Avila, C., Moltó, J. The sensitivity to punishment and sensitivity to reward questionnaire (SPSRQ) as a measure of Gray's anxiety and impulsivity dimensions. Treadway, M. T., Buckholtz, J. W., Schwartzman, A. N., Lambert, W. E. & Zald, D. H. Worth the ‘EEfRT'? The effort expenditure for rewards task as an objective measure of motivation and anhedonia. Knutson, B., Westdorp, A., Kaiser, E. & Hommer, D. FMRI visualization of brain activity during a monetary incentive delay task. Kirschner, M. et al. Deficits in context-dependent adaptive coding of reward in schizophrenia. Murty, V. P. et al. Resting state networks distinguish human ventral tegmental area from substantia nigra. Koush, Y. et al. OpenNFT: an open-source Python/Matlab framework for real-time fMRI neurofeedback training based on activity, connectivity and multivariate pattern analysis. The statistical analysis of fMRI data. Hinds, O. et al. Computing moment-to-moment BOLD activation for real-time neurofeedback. Koush, Y., Zvyagintsev, M., Dyck, M., Mathiak, K. A. & Mathiak, K. Signal quality and Bayesian signal processing in neurofeedback based on real-time fMRI. Bagarinao, E., Matsuo, K., Nakai, T. & Sato, S. Estimation of general linear model coefficients for real-time application. McCrae, R. R. & Costa, P. T. Empirical and theoretical status of the five-factor model of personality traits. SAGE Handb. Theory Assess. Esteban, O. et al. fMRIPrep: a robust preprocessing pipeline for functional MRI. Jenkinson, M., Beckmann, C. F., Behrens, T. E., Woolrich, M. W. & Smith, S. M. Fsl. Oldham, S. et al. The anticipation and outcome phases of reward and loss processing: a neuroimaging meta-analysis of the monetary incentive delay task. Brain Mapp. Woolrich, M. Robust group analysis using outlier inference. Beckmann, C. F., Jenkinson, M. & Smith, S. M. General multilevel linear modeling for group analysis in fMRI. Woolrich, M. W., Behrens, T. E. J., Beckmann, C. F., Jenkinson, M. & Smith, S. M. Multilevel linear modelling for FMRI group analysis using Bayesian inference. R Core Team, R. R: A Language and Environment for Statistical Computing (R Foundation for Statistical Computing, 2020). Chikhi, S., Matton, N., Sanna, M. & Blanchet, S. Mental strategies and resting state EEG: effect on high alpha amplitude modulation by neurofeedback in healthy young adults. Kober, S., Witte, M., Ninaus, M., Neuper, C. & Wood, G. Learning to modulate one's own brain activity: the effect of spontaneous mental strategies. McLaren, D. G., Ries, M. L., Xu, G. & Johnson, S. C. A generalized form of context-dependent psychophysiological interactions (gPPI): a comparison to standard approaches. Sescousse, G., Caldú, X., Segura, B. Processing of primary and secondary rewards: a quantitative meta-analysis and review of human functional neuroimaging studies. Virtanen, P. et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Kühn, S. & Gallinat, J. Segregating cognitive functions within hippocampal formation: a quantitative meta-analysis on spatial navigation and episodic memory. Brain Mapp. Arsalidou, M. & Taylor, M. J. Meta-analyses of brain areas needed for numbers and calculations. Hétu, S. et al. The neural network of motor imagery: an ALE meta-analysis. McNorgan, C. A meta-analytic review of multisensory imagery identifies the neural correlates of modality-specific and modality-general imagery. Download references This study was funded by Joy Ventures foundation and by KAMIN program in the Israeli ministry of innovation (grant 2029678; A.R. and Teva Bio-Innovation Forum (N.L.). would like to thank the AMRF Adelson Family Foundation support. We thank Y. Benjamini for his statistical advice, N. Noy for her assistance in creating the graphical illustrations depicted in Fig. 1, A. Kuzli and M. Szwarcwort for their assistance in measuring plasma antibody titers, S. Schwartzbaum for her help with editing, and R. Cohen for his inspiration and encouragement for pursuing this animal-to-human translation effort. These authors contributed equally: Nitzan Lubianiker, Tamar Koren, Asya Rolls, Talma Hendler. School of Psychological Sciences, Tel Aviv University, Tel Aviv, Israel Nitzan Lubianiker, Itamar Jalon & Talma Hendler Department of Psychology, Yale University, New Haven, CT, USA Nitzan Lubianiker Princeton Neuroscience Institute, Princeton University, Princeton, NJ, USA Nitzan Lubianiker & Itamar Jalon Department of Pathology, Tel Aviv Sourasky Medical Center, in affiliation with the Gray Faculty of Medical and Health Sciences at Tel Aviv University, Tel Aviv, Israel Tamar Koren, Margarita Sirotkin, Hilla Azulay-Debby & Asya Rolls School of Neuroscience, Biochemistry and Biophysics, George S. Wise Faculty of Life Sciences, Tel Aviv University, Tel Aviv, Israel Tamar Koren, Margarita Sirotkin, Hilla Azulay-Debby & Asya Rolls Sagol Brain Institute, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel Meshi Djerasi, Neomi Singer, Avigail Lerner, Haggai Sharon & Talma Hendler Sagol School of Neuroscience, Tel Aviv University, Tel Aviv, Israel Meshi Djerasi, Neomi Singer, Avigail Lerner & Talma Hendler Gray Faculty of Medical and Health Sciences, Tel Aviv University, Tel Aviv, Israel Roi Sar-el, Haggai Sharon & Talma Hendler The Tel-Aviv University Center for AI and Data Science, Tel Aviv, Israel Rappaport Faculty of Medicine, Technion, Israel Institute of Technology, Haifa, Israel Asya Rolls Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar conceptualized the study and designed the experiments. and T.H conceptualized the MSQ questionnaire. developed the randomized-network control condition and the mental strategies characterization protocol. developed online fMRI-NF analysis pipelines. collected the behavioral, neural, and mental data. collected blood samples and vaccinated participants. analyzed the behavioral and neuroimaging data. collected and analyzed the immunological data. provided statistical advice. secured funding and supervised the study. wrote the paper. All authors edited the paper. Correspondence to Nitzan Lubianiker, Asya Rolls or Talma Hendler. is the Chief Medical Scientist and Chair of the advisory board in GrayMatters Health. have a filed patent related to the topic of this paper in the United States Patent and Trademark Office (application number: 17/435.906; Title of Invention: Neurofeedback and Induction of an Immune Response). The other authors declare no competing interests. Nature Medicine thanks Ulrike Bingel, Marta Peciña and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Jerome Staal, in collaboration with the Nature Medicine team. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Whole-brain results for the high vs neutral reward anticipation contrast during the Monetary Incentive Delay (MID) task. The overlaid predefined meta-analytic masks of the mesolimbic system, taken from refs. 57,66, are shown in white. Abbreviations: Nucleus Accumbens (Nac); Ventral Tegmental Area (VTA). Statistical threshold was set at a cluster-level p(FWE) < .05. b. ROIs of the reward mesolimbic network. The offline functional localization of the reward mesolimbic regions. For the VTA, the predefined anatomical mask57 was intersected with the high>neutral reward anticipation maps presented in A. For the bilateral Nac, a 5 mm sphere was created around the peak activation voxel within the 8 mm predefined Nac masks (defined based on reward anticipation meta-analysis66). In randomized ROI-NF, each participant is randomly assigned to one of K subgroups of different functionally defined neural targets (colored in pink, orange, and yellow). While in each sub-group, a specific neural target is modulated along with the general task processes (colored in green – control, learning and reward processes), group-level modulations (right panel) include only the averaged (non-specific) general processes common to all subgroups (that is the target process specific to each subgroup is averaged over the whole group). Conversely, in a typical experimental NF group, all participants modulate both the general and target processes (colored in green and blue, respectively). Therefore, group-level modulations include both the general processes and the reoccurring target process modulations. Consequently, differences in outcome effects (such as immune function) between groups can be attributed solely to the target-specific effects (Figure adapted from ref. b. Neural networks of subgroups. Targets of the randomized ROI-NF control condition. All target networks were selected from meta-analyses of fMRI studies on specific functional processes. Spatial navigation network76: right medial temporal lobe: 26, -35, -11; left medial temporal lobe: -26, -47, -9; left posterior cingulate: -15, -59, 19; T2. Arithmetic processing network77: right SPL: 29, -66, 49; left dlpfc: -45, 32, 29; left precuneus: -28, -71, 33; T3; Motor imagery network78: right cerebellum: 32, -62, -28; left cerebellum: -32, -56, -30; left precentral gyrus: -26, -2, 58; T4. Auditory imagery network79: right STG:64, -30, 9; left IFG: -48, 24, −5; left precentral gyrus: -52 1 47. Following practice, participants fill out the MSQ-NF for each strategy applied during the session, based on their logged choices during training. “Other” strategies, as well as those with predefined names, are verbally described, and then each strategy is labeled across the MSQ space. a. Upregulate>Rest contrast, showing differing regulation effects per group (reward ml NF – light blue, randomized ROI NF – pink). Feedback contrast. c. Choice contrast. Abbreviations: right Insula (r Ins); subgenual ACC (sgACC); Thalamus (Thal), ventromedial prefrontal cortex (vmpfc). The VTA and bilateral Nac ROIs are overlaid in white. Supplementary Results, Supplementary protocol 1, and Supplementary Table 1, and Supplementary Figures 1-3. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. Reprints and permissions Lubianiker, N., Koren, T., Djerasi, M. et al. Upregulation of reward mesolimbic activity and immune response to vaccination: a randomized controlled trial. Nat Med (2026). Download citation Received: 25 June 2024 Accepted: 20 November 2025 Published: 19 January 2026 Version of record: 19 January 2026 DOI: https://doi.org/10.1038/s41591-025-04140-5 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Nature Medicine (Nat Med) ISSN 1546-170X (online) ISSN 1078-8956 (print) © 2026 Springer Nature Limited Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

A tool-using cow is challenging what we know about farm animal intelligence A pet cow named Veronika uses tools in a surprisingly sophisticated way—possibly because she has been allowed to live her best life In news that is sure to delight fans of a certain Gary Larson cartoon turned meme about the limitations of bovine cognition, cow tools are real. Larson's 1982 comic for his series The Far Side showed a cow standing behind a table bearing an array of oddly shaped objects. The finding adds a new species to the growing list of creatures that have been found to use external objects to achieve a goal and suggests that society has been underestimating the minds of farm animals. The story begins more than a decade ago with Witgar Wiegele, an organic farmer and traditional baker in the small Austrian town of Nötsch im Gailtal. Wiegele first observed that his family's pet Swiss Brown cow, Veronika, would sometimes pick up sticks and use them to scratch herself, presumably to alleviate skin irritation from insects. When cognitive biologist Alice Auersperg of the University of Veterinary Medicine, Vienna, saw a video recording of Veronika's behavior, “it was immediately clear that this was not accidental,” she said in a statement. If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today. “Veronika is very friendly,” says Osuna-Mascaró, who spent the summer observing her. “She also has a close bond with Witgar,” he notes. “Not only does Witgar prepare and sell bread, he also distributes it around the area. It was interesting to see Veronika watching every passing car with interest and trying to guess if the driver was Witgar. The researchers analyzed how Veronika used one particular tool—a deck brush—to scratch herself. She also scrubbed more vigorously on tougher parts of her skin and used gentle pushes on her delicate parts. Among non-human species, this kind of tool use has only been consistently documented in chimpanzees. “We don't believe that Veronika is the Einstein of cows,” Osuna-Mascaró says. Together with anecdotal reports of tool use in cattle from South Asia, the results of the new study hint that the capacity for complex problem-solving behaviors, including tool use, might have ancient evolutionary roots but that such behaviors emerge only when conditions are favorable. Nötsch im Gailtal is “the most idyllic place imaginable for an Austrian cow, like straight out of The Sound of Music,” Osuna-Mascaró says. He says the family contributed to Veronika's tool use by “providing the special conditions that enabled Veronika to express herself.” Although she learned to use tools by herself, starting with branches that had fallen from trees, Wiegele later furnished her with sticks and rakes that allowed her to perfect her scratching techniques. Most livestock animals, in contrast, live much shorter lives and spend their time in impoverished settings such as factory farms without access to objects that they can manipulate. I applaud the authors, as well as Veronika!” says primatologist Jill Pruetz of Texas State University, who was not involved in the new research. Pruetz studies how environmental factors influence the behavior of tool-using chimpanzees. She also has two companion cows of her own, Claire and Edith. “I am not completely surprised that cattle can use tools—after living in close proximity to my two cows for about seven years now, I have a lot more respect for bovine intelligence!” Pruetz says. “What strikes me about Veronika's tool use is the precision with which she can manipulate the tool as well as switch its ends to target specific areas." “There are around 1.5 billion heads of cattle in the world, and humans have lived with them for at least 10,000 years. It's shocking that we're only discovering this now,” Osuna-Mascaró says. Kate Wong is an award-winning science writer and senior editor for features at Scientific American, where she has focused on evolution, ecology, anthropology, archaeology, paleontology and animal behavior. Her reporting has taken her to caves in France and Croatia that Neandertals once called home to the shores of Kenya's Lake Turkana in search of the oldest stone tools in the world, as well as to Madagascar on an expedition to unearth ancient mammals and dinosaurs, the icy waters of Antarctica, where humpback whales feast on krill, and a “Big Day” race around the state of Connecticut to find as many bird species as possible in 24 hours. Wong is co-author, with Donald Johanson, of Lucy's Legacy: The Quest for Human Origins. If you enjoyed this article, I'd like to ask for your support. Scientific American has served as an advocate for science and industry for 180 years, and right now may be the most critical moment in that two-century history. If you subscribe to Scientific American, you help ensure that our coverage is centered on meaningful research and discovery; that we have the resources to report on the decisions that threaten labs across the U.S.; and that we support both budding and working scientists at a time when the value of science itself too often goes unrecognized. In return, you get essential news, captivating podcasts, brilliant infographics, can't-miss newsletters, must-watch videos, challenging games, and the science world's best writing and reporting. There has never been a more important time for us to stand up and show why science matters.

You are using a browser version with limited support for CSS. In the weeks leading up to September 1891, mathematician Georg Cantor prepared an ambush. For years he had sparred — philosophically, mathematically and emotionally — with his formidable rival Leopold Kronecker, one of Germany's most influential mathematicians. Stirring biopic of the first woman to win top maths prize Stirring biopic of the first woman to win top maths prize But Cantor had a proof that he hoped would confound his competitor. Armed with an innovative method, now known as Cantor's diagonal argument, he could demonstrate that some infinities are larger than others and he planned to confront Kronecker with it in public at the inaugural meeting of the German Mathematical Society in Halle. Weeks before the meeting, Kronecker's wife was fatally injured in a climbing accident, preventing him from going to Halle, and Kronecker himself died that December. This tragedy — at once poignant, anticlimactic and painfully human — is one of many vivid digressions in journalist Jason Socrates Bardi's The Great Math War. Was maths grounded in intuition — the mental constructions that people can build step by step — or in formal symbols and rules that are independent of human insight? The book's central argument — that maths is a human endeavour, shaped by human frailty as much as by formal proof — is made convincingly throughout. Hilbert championed formalism: the idea that maths should be reduced to axioms and rules, with meaning set aside in favour of consistency. In 1928, Hilbert, who was a chief editor of the leading journal Mathematische Annalen, tried to remove Brouwer from its editorial board. Brouwer refused to resign and was ousted — through dissolution of the entire board. Albert Einstein, a fellow Annalen chief editor, dismissed the whole affair as an overblown ‘frog and mouse war'. For many readers who have encountered maths only as a technical tool, the heightened passions behind such debates might be hard to fathom. Yet, when Cantor was being denounced by his contemporaries, including Kronecker, as a charlatan and accused of corrupting the youth with his controversial ideas about infinity, the future of maths truly seemed at stake. He writes with breathless enthusiasm, occasionally slipping into staccato bursts — “Bereft. He brings comic energy to his narrative, for instance, by introducing ancient Greek mathematician Euclid as the “reigning heavyweight champion of math authors” until 1900. An imperialist turned peace activist and advocate of free love, Russell discovered a troubling paradox in maths, similar in its logical form to several others, including the ‘liar paradox': if the sentence ‘this statement is a lie', for instance, is true, then it means the statement is a lie — and therefore not true — leading to a logical inconsistency. Stirring biopic of the first woman to win top maths prize AI tools boost individual scientists but could limit research as a whole Training large language models on narrow tasks can lead to broad misalignment During the course of my PhD, I've been relearning how to rest Job Title:    Senior Editor, Discover Journals (Medicine, Public Health, and Clinically Oriented Research) Location:   Beijing/Shanghai/Nanjing/Pu... Join HZAU's global faculty team to advance research with competitive benefits. The Taikang Center at Wuhan University seeks exceptional faculty for multiple open-rank positions, providing top-tier support in key biomedical field Stirring biopic of the first woman to win top maths prize An essential round-up of science news, opinion and analysis, delivered to your inbox every weekday. Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). Nevertheless, a century still seemed a remote and unattainable score, something that only the most distinguished professors were likely to achieve before they retired. I have always felt the pressure to publish more. In 2005, my head of department often exalted a professor who published around 40 papers a year; those of us working in public health were all encouraged to do the same. I've decided to push back on the pressure to publish by making a rule for myself: I will no longer publish more than seven papers per year. I've published a median of 15 papers a year for the past five years, so that means I will have to halve my output. Seven is my own threshold, and will have no relevance to most other researchers, especially those in other fields or at different career stages. I am not going to spend less time on research, or contribute to 15 papers per year and then select the best 7, so this change will double the amount of time I spend on each paper. Publication numbers are ballooning, with more than 1.7 million indexed articles appearing on the PubMed database for 2024, compared with around 1.2 million for 2014. This inflation was recently described as “unsustainable” by Cambridge University Press, UK, which called for “radical change”. Without adequate scrutiny, the overall quality of research seems to be degrading, with a recent explosion in the number of low-quality papers passing peer review3. As a professor with a permanent contract, it's much easier for me to make radical changes than it is for early-career researchers, who are competing to win jobs and funding. Too many career decisions are based on CV length rather than quality, which maintains the pressure on researchers to publish as many papers as possible. I'm not the first to suggest that less is more in publishing. Alas, speed is still more prized than rigour, because those who publish first get more kudos than do those who publish carefully. Subscribe to Nature Briefing: Careers, an unmissable free weekly round-up of help and advice for working scientists. I am a member of the Association for Interdisciplinary Meta-Research and Open Science (AIMOS) whose aims include increasing research quality. Publishing nightmare: a researcher's quest to keep his own work from being plagiarized During the course of my PhD, I've been relearning how to rest Do you have a side hustle alongside your PhD studies? During the course of my PhD, I've been relearning how to rest Do you have a side hustle alongside your PhD studies? AI tools boost individual scientists but could limit research as a whole Job Title:    Senior Editor, Discover Journals (Medicine, Public Health, and Clinically Oriented Research) Location:   Beijing/Shanghai/Nanjing/Pu... Join HZAU's global faculty team to advance research with competitive benefits. The Taikang Center at Wuhan University seeks exceptional faculty for multiple open-rank positions, providing top-tier support in key biomedical field Seeking exceptional Senior/Junior PIs, Postdocs, and Core Specialists globally year-round Publishing nightmare: a researcher's quest to keep his own work from being plagiarized An essential round-up of science news, opinion and analysis, delivered to your inbox every weekday. Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Roberto Patricio Korzeniewicz is sociologist at the University of Maryland, College Park. The Great Global Transformation: National Market Liberalism in a Multipolar World Branko Milanovic Allen Lane (2025) Economist Branko Milanovic maps out those changes, and what they might herald, in his latest book. By contrast, relative shares of world income stagnated or were thought to have declined for wealthy nations' middle and working classes, including in the United States. This trend captures effectively what has become part of public discourse in the West: globalization has left behind people in the middle and working classes in rich nations, demonstrating that it is “an illusion to believe that what is good for the world must be good for everybody”. China leads research in 90% of crucial technologies — a dramatic shift this century China leads research in 90% of crucial technologies — a dramatic shift this century At the same time, opportunities for upward social mobility in wealthy nations have been transformed by an increasing overlap of people who are rich in terms of receiving income both from capital (through investments and assets) and from labour (salaries). Although there are national differences, these countries have in common several of the elements that were previously encapsulated by neoliberalism and globalization. They are also committed to disciplining wealthy people and regulating free trade through various degrees of state control. AI is transforming the economy — understanding its impact requires both data and imagination AI is transforming the economy — understanding its impact requires both data and imagination AI is transforming the economy — understanding its impact requires both data and imagination China leads research in 90% of crucial technologies — a dramatic shift this century Why the global economy is more uncertain than ever, and what to do about it The academic community failed Wikipedia for 25 years — now it might fail us Job Title:    Senior Editor, Discover Journals (Medicine, Public Health, and Clinically Oriented Research) Location:   Beijing/Shanghai/Nanjing/Pu... The Taikang Center at Wuhan University seeks exceptional faculty for multiple open-rank positions, providing top-tier support in key biomedical field AI is transforming the economy — understanding its impact requires both data and imagination China leads research in 90% of crucial technologies — a dramatic shift this century Why the global economy is more uncertain than ever, and what to do about it An essential round-up of science news, opinion and analysis, delivered to your inbox every weekday. Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

EPA's pollution rule change worries experts, cancer survival hits milestone, and astronauts evacuate the ISS By Kendra Pierre-Louis, Andrea Thompson, Sushmita Pathak & Alex Sugiura You're listening to our weekly science news roundup. First up, earlier this month the U.S. Environmental Protection Agency published a new rule signaling a major change to the way it accounts for the impact of certain air pollutants on human health. Andrea Thompson, SciAm's senior editor for life sciences, is here to give us a clearer understanding of what the agency is doing and what that means for the air we all breathe. If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today. Pierre-Louis: So my understanding is that, in general, the EPA uses a type of cost-benefit analysis to understand the impact of an environmental regulation. Can you talk a bit about how, prior to this rule change, the EPA measured these costs and benefits? Thompson: Yeah, so in general the science for how you sort of calculate how much a regulation will cost, the sort of economic benefit you'll get from it has been pretty well established over several decades. So for, at least for the benefits part, it starts with health studies that compare areas where, say, a certain pollutant is high and areas where it's low, and look at the differences in hospitalizations and premature deaths and other health indicators, and controlling for other factors that may influence those, and come up with a model that you can use to see, “Okay, if this pollutant goes up or down, how much do those health indicators change?” And then that's married with economic studies that sort of look at what's called the “value of, of a statistical life”—so these aren't a moral judgment of how much a life is worth [Laughs]; it's a statistical one. And that sort of gets married together to figure out, “Okay, here's what the economic benefits will be if we regulate this pollutant.” Thompson: So it's a little unclear, as the EPA hasn't been very specific on this point. From the language cited in a New York Times story that came out they're at least not calculating a, you know, a sort of dollar value for the health benefits. And they said that they're not going to calculate that for PM2.5, which are little particles that can penetrate very deeply in the lungs, into the bloodstream; are linked to a bunch of health effects, from asthma to cancers to low-birth-weight babies. Pierre-Louis: Right, these are the kinds of things that we encounter when we burn gasoline for fuel or during the wildfire smoke events. Pierre-Louis: So in the old days, let's say, I had a plant, and it was polluting, and it cost me, I don't know, $100 to put a scrubber on my plant so it had less particle pollution, and the government could say, “Well, yeah, it's gonna cost you $100, but it's gonna save $1,000 in human health costs.” But now they're still looking at the $100, but they no longer have that $1,000 to compare it to. Pierre-Louis: And when the news first broke of this change it seemed like many health advocates, their response was, like, “This is likely going to lead to higher levels of these pollutants.” Have you heard something similar? Thompson: Yeah, so from the experts I've talked to, you know, this would mean that anything that falls under the umbrella of regulations where this is the new policy, you're going to probably have higher levels of air pollutants than you would have otherwise because it's hamstringing this critical tool to figure out, you know, if regulations are going to be worthwhile. Pierre-Louis: If changing the rules in this way is likely to lead to more air pollution, why do it? Pierre-Louis: I know some people listening at home might be wondering: Is there something their state can do to impose stricter air pollution rules than the EPA can, for example? Thompson: Right, so states often have particular environmental rules. The problem with something like air pollution is that you could have a polluting plant in, say, Pennsylvania, and those pollutants are going to be blown over into New Jersey and New York. Turning to some news on cancer, 70 percent of cancer patients now survive at least five years after diagnosis, according to the most recent annual report of the American Cancer Society. Rebecca Siegel, the organization's senior scientific director for surveillance research, said in a statement, “This stunning victory is largely the result of decades of cancer research that provided clinicians with the tools to treat the disease more effectively, turning many cancers from a death sentence into a chronic disease.” The five-year survival rate for myeloma, a kind of blood cancer, jumped from 32 percent in the mid-'90s to 62 percent from 2015 to 2021. Similarly, five-year survival numbers for regional-stage lung cancer, which is typically stage 3, rose from 20 percent to 37 percent over the same time period. The researchers cited improved screening and cancer treatments, as well as a decline in smoking, for these positive outcomes. But the authors also cautioned that recent shifts in federal policy could undo this progress. The report concluded that “pending federal cuts to health insurance and cancer research will inevitably reduce access to life‐saving drugs and halt progress at a time when incidence is rising for many common cancers.” Speaking of illnesses, NASA astronauts Mike Fincke and Zena Cardman, Russian cosmonaut Oleg Platonov and Japanese astronaut Kimiya Yui, also known as Crew-11, splashed down off the coast of California on Thursday following a medical evacuation. Crew-11 had been expected to stay on the International Space Station through mid-February, but NASA ordered the departure after one of the astronauts developed what the agency's administrator called “a serious medical condition.” Due to medical privacy rules NASA has not revealed which astronaut fell ill or what condition they developed. Lessons from this evacuation could help prepare for upcoming human spaceflight missions, including Artemis II. Researchers have known for a while that same-sex sexual behavior is common in animals. But in a study published last Monday a team of scientists offered more insight into the potential evolutionary underpinnings of this behavior in primates. Imperial College London researchers looked at data on 491 non-human primate species and found same-sex sexual behaviors in 59 of them. The scientists discovered that same-sex sexual interaction was more likely for species with certain characteristics. For example, primates who lived in drier places susceptible to greater food scarcity and predation pressure. Tune in on Wednesday, when we'll take a deep dive into the scientific quest to define consciousness. Record a voice memo on your phone or computer and send it over to ScienceQuickly@sciam.com. Be sure to include your name and where you're from. Science Quickly is produced by me, Kendra Pierre-Louis, along with Fonda Mwangi, Sushmita Pathak and Jeff DelViscio. Shayna Posses and Aaron Shattuck fact-check our show. Our theme music was composed by Dominic Smith. Subscribe to Scientific American for more up-to-date and in-depth science news. She has worked for Gimlet, Bloomberg News and Popular Science. Andrea Thompson is senior desk editor for life science at Scientific American, covering the environment, energy and earth sciences. She has been covering these issues for nearly two decades. Prior to joining Scientific American, she was a senior writer covering climate science at Climate Central and a reporter and editor at Live Science, where she primarily covered earth science and the environment. She has moderated panels, including as part of the United Nations Sustainable Development Media Zone, and appeared in radio and television interviews on major networks. She holds a graduate degree in science, health and environmental reporting from New York University, as well as a B.S. in atmospheric chemistry from the Georgia Institute of Technology. She previously worked at NPR and was a regular contributor to The World from PRX and The Christian Science Monitor. He has worked on projects for Bloomberg, Axios, Crooked Media and Spotify, among others. If you enjoyed this article, I'd like to ask for your support. Scientific American has served as an advocate for science and industry for 180 years, and right now may be the most critical moment in that two-century history. I hope it does that for you, too. If you subscribe to Scientific American, you help ensure that our coverage is centered on meaningful research and discovery; that we have the resources to report on the decisions that threaten labs across the U.S.; and that we support both budding and working scientists at a time when the value of science itself too often goes unrecognized. In return, you get essential news, captivating podcasts, brilliant infographics, can't-miss newsletters, must-watch videos, challenging games, and the science world's best writing and reporting. There has never been a more important time for us to stand up and show why science matters. I hope you'll support us in that mission. Subscribe to Scientific American to learn and share the most exciting discoveries, innovations and ideas shaping our world today.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. (2026)Cite this article The rapid growth of Internet of Things applications has substantially increased the number of connected sensors and data volume, yet conventional digital conversion and transmission systems impose high energy and latency costs. Here we develop a neuromorphic sensing system integrating a flexible piezoelectric haptic sensor array, event-triggered preprocessing circuitry and a memristive system on a chip. The circuitry transforms transient voltage spikes from sensor pixels into decaying voltage waveforms, generating a time surface for event-based analogue in-memory computing within the chip. Our system achieves 87%–92% recognition accuracy for patterns written on the sensor array and reduces the energy-delay product during inference compared with conventional digital platforms. These results highlight the potential of the memristive system on a chip for energy-efficient, low-latency edge processing of analogue sensor data, advancing intelligent sensing technologies. This is a preview of subscription content, access via your institution Subscribe to this journal Receive 12 digital issues and online access to articles $119.00 per year only $9.92 per issue Buy this article Prices may be subject to local taxes which are calculated during checkout The data that support the plots within this article, as well as other findings of this study, are available via Code Ocean at https://codeocean.com/capsule/4914834/tree/v1. The code that supports the results shown in this article is available via Code Ocean at https://codeocean.com/capsule/4914834/tree/v1. Code that supports the operation of the integrated chip for the demonstrated application is available from the corresponding author upon reasonable request. Atzori, L., Iera, A. & Morabito, G. The Internet of Things: a survey. Shi, W., Cao, J., Zhang, Q., Li, Y. & Xu, L. Edge computing: vision and challenges. IEEE Internet Things J. Zhou, F. & Chai, Y. Near-sensor and in-sensor computing. Hennessy, J. L. & Patterson, D. A. Computer Architecture: A Quantitative Approach (Elsevier, 2011). Schuman, C. D. et al. A survey of neuromorphic computing and neural networks in hardware. Preprint at https://arxiv.org/abs/1705.06963 (2017). Gallego, G. et al. Event-based vision: a survey. IEEE Trans. Brandli, C., Berner, R., Yang, M., Liu, S. C. & Delbruck, T. A. 240 × 180 130 dB 3 μs latency global shutter spatiotemporal vision sensor. IEEE J. Solid-State Circuits 49, 2333–2341 (2014). Posch, C., Matolin, D. & Wohlgenannt, R. A QVGA 143dB dynamic range asynchronous address-event PWM dynamic image sensor with lossless pixel-level video compression. In 2010 IEEE International Solid-State Circuits Conference 400–401 (IEEE, 2019). Lichtsteiner, P. & Delbruck, T. A 64×64 AER logarithmic temporal derivative silicon retina. In 2005 PhD Research in Microelectronics and Electronics – Proceedings of the Conference II 406–409 (IEEE, 2005). Chen, S. & Guo, M. Live demonstration: CeleX-V: a 1M pixel multi-mode event-based sensor. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops 1682–1683 (IEEE, 2019). Liao, F. et al. Bioinspired in-sensor visual adaptation for accurate perception. Zhou, F. et al. Optoelectronic resistive random access memory for neuromorphic vision sensors. Sun, L. et al. In-sensor reservoir computing for language learning via two-dimensional memristors. Lee, D. et al. In-sensor image memorization and encoding via optical neurons for bio-stimulus domain reduction toward visual cognitive processing. Li, Z. et al. Crossmodal sensory neurons based on high-performance flexible memristors for human-machine in-sensor computing system. Mennel, L. et al. Ultrafast machine vision with 2D material neural network image sensors. Jang, H. et al. In-sensor optoelectronic computing using electrostatically doped silicon. Zhou, Y. et al. Computational event-driven vision sensors for in-sensor spiking neural networks. Zhou, Y. et al. A 2T2R1C vision cell with 140 dB dynamic range and event-driven characteristics for in-sensor spiking neural network. In 2022 International Electron Devices Meeting 3141–3144 (IEEE, 2022). Prezioso, M. et al. Training and operation of an integrated neuromorphic network based on metal-oxide memristors. Yao, P. et al. Fully hardware-implemented memristor convolutional neural network. Pi, S. et al. Memristor crossbar arrays with 6-nm half-pitch and 2-nm critical dimension. Cai, F. et al. A fully integrated reprogrammable memristor–CMOS system for efficient multiply–accumulate operations. Hu, M. et al. Dot-product engine for neuromorphic computing: programming 1T1M crossbar to accelerate matrix–vector multiplication. In Proc. 53rd Annual Design Automation Conference 1–6 (IEEE, 2016). & Yang, J. J. Memristive crossbar arrays for brain-inspired computing. Yoon, J. H. et al. An artificial nociceptor based on a diffusive memristor. Jiang, H. et al. Sub-10 nm Ta channel responsible for superior performance of a HfO2 memristor. Rao, M. et al. Thousands of conductance levels in memristors integrated on CMOS. Xia, Q. et al. Memristor–CMOS hybrid integrated circuits for reconfigurable logic. Nano Lett. Li, C. et al. Analogue signal and image processing with large memristor crossbars. Wang, Z. et al. Reinforcement learning with analogue memristor arrays. Li, C. et al. Long short-term memory networks in memristor crossbar arrays. Li, C. et al. Efficient and self-adaptive in-situ learning in multilayer memristor neural networks. Li, C. et al. In-memory computing with memristor arrays. In 2018 IEEE 10th International Memory Workshop 1–4 (IEEE, 2018); https://doi.org/10.1109/IMW.2018.8388838 Dang, B. et al. Reconfigurable in-sensor processing based on a multi-phototransistor–one-memristor array. Huang, H. et al. Fully integrated multi-mode optoelectronic memristor array for diversified in-sensor computing. Lagorce, X., Orchard, G., Galluppi, F., Shi, B. E. & Benosman, R. B. HOTS: a hierarchy of event-based time-surfaces for pattern recognition. IEEE Trans. Sironi, A., Brambilla, M., Bourdis, N., Lagorce, X. & Benosman, R. HATS: histograms of averaged time surfaces for robust event-based object classification. In Proc. IEEE Computer Society Conference on Computer Vision and Pattern Recognition 1731–1740 (IEEE, 2018). Song, W. et al. Programming memristor arrays with arbitrarily high precision for analog computing. Wang, Z. et al. Multi-diseases detection with memristive system on chip. Huang, Y. et al. Radiofrequency signal processing with a memristive system-on-a-chip. Lloyd, S. P. Least squares quantization in PCM. IEEE Trans. Some methods for classification and analysis of multivariate observations. In Proc. 5th Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Statistics Vol. 5.1 (eds Le Cam, L. M. & Neyman, J.) Akopyan, F. et al. TrueNorth: design and tool flow of a 65 mW 1 million neuron programmable neurosynaptic chip. IEEE Trans. Aided Des. Pehle, C. et al. The BrainScaleS-2 accelerated neuromorphic system with hybrid plasticity. Milano, G. et al. In materia reservoir computing with a fully memristive architecture based on self-organizing nanowire networks. Research in the USA is mainly sponsored by the National Science Foundation (CCF, grant no. ), and in part by the Army Research Laboratory (grant no. and the Office of Naval Research (grant no. ); research in Finland is mainly sponsored by the Research Council of Finland (grant no. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Laboratory or the US Government. The US Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. Department of Electrical & Computer Engineering, University of Massachusetts Amherst, Amherst, MA, USA Wuyu Zhao, Yi Huang, Alireza Jaberi Rad & Qiangfei Xia Faculty of Information Technology and Communication Sciences, Tampere University, Tampere, Finland Amit Tewari & Sayani Majumdar Millburn High School, Millburn, NJ, USA TetraMem Inc., San Jose, CA, USA Ning Ge, J. Joshua Yang, Miao Hu & Qiangfei Xia Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, USA J. Joshua Yang Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar conceived and led the project. conducted work on the preprocessing circuitry and the SoC-based computing, A.T. and S.M. made the piezoelectric sensors. contributed to the evaluation kit (hardware and software). wrote the article. All authors edited the paper before submission. Correspondence to Qiangfei Xia. are cofounders and paid consultants of TetraMem. The other authors declare no competing interests. Nature Sensors thanks Laura Begon-Lours and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Combined transient response of single pads across a pressure range (1.4-59.7 kPa). Transient response at 1.4 kPa. c. High-resolution transient waveform from 2 to 4 s. d Sensor sensitivity (voltage vs. pressure). e. Response time calculation. f. Sensor response test after bending (10.4% strain and 22.2% strain). g. Photograph of the bending test system set-up and when strain = 0% (Original Size:9.6 cm) and h. when strain = 10.4% and bending cycle = 1030 (Vertical change in size: 1 cm, Radius: 5.9 cm) and i. when strain = 10.4% and bending cycle = 1010. (Vertical change in size: 2 cm, radius: 4.2 cm). The combination of the sensor array and the connection circuitry converts the time-related trail information into the voltage amplitude. These pixels with different voltage amplitudes will be treated as images, and we train a CNN for that application. Different kernel information and the weight information in the fully connected layer are programmed into the memristive array for inference. b. CNN structure for this 3-letter recognition task c. the block division of the memristive array. d. Conductance map of the target and actual weight mapping of the CNN. e. Accuracy of the 3-letter recognition task. It reaches 100 percent accuracy for both software and hardware results. For this task we use 2 convolutional layers and 1 fully connected layer structure with 2 max-pooling layers for the reduction of the dimensions. Block division of a single 248×256 memristive array for this CNN. c. Target and actual conductance map comparison for the weight mapping of both the convolutional and fully connected layers. d. Accuracy of the writing direction in the digit's recognition task. We obtain 100% and 99.83% accuracy from the neural network-based computing method from software and hardware results separately. Supplementary Videos 1 and 2 captions, Table 1, Figs. 1–6, Notes 1–6 and References. Sensor response under manual pressing. The video demonstrates the testing procedure used to obtain the piezoelectric sensor array's response under repeated manual pressing. Controlled press test using a MARK-10 pressure gauge. The video shows the procedure for obtaining the piezoelectric sensor array's response under precisely controlled pressing conditions. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. Reprints and permissions Zhao, W., Huang, Y., Tewari, A. et al. Event-based neuromorphic sensing system with flexible haptic sensors and a memristive system on a chip. Download citation Received: 17 June 2025 Version of record: 19 January 2026 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Nature Sensors ISSN 3059-4499 (online) © 2026 Springer Nature Limited Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. The lack of biomarkers to identify individuals at risk of asthma exacerbations remains a significant limitation to improving patient outcomes. To address this need, we analyze data from three asthma cohorts, combining up to 25 years of electronic medical records with sequential metabolomics studies, to develop and replicate a predictive model for asthma exacerbation risk. We identify asthma-associated biochemical pathways via global circulatory metabolomics and then apply targeted mass spectrometry methods to quantify selected steroids, sphingolipids, and microbial-derived metabolites. The sphingolipid-to-steroid ratios robustly associate with 5-year exacerbation risk (discovery p value = 1.63×10⁻26-0.029; replication p value = 1.89×10⁻36-0.033). Based upon these findings, we derive and replicate a simple 5-year predictive model of asthma exacerbations using 21 sphingolipid-to-steroid ratios that outperforms current clinical measures (discovery AUC = 0.90; replication AUC = 0.89). These findings underscore the value of metabolomic profiling to develop a practical, cost-effective clinical assay for asthma exacerbation risk that may improve patient care. Asthma exacerbations are a major health care burden and cause of disease morbidity, leading to progressive loss of lung function, airway remodeling, and enhanced disease severity1,2. The heterogeneity of asthma complicates identifying individuals at high risk for future exacerbations3,4, and there are no current clinical tests or biomarkers that effectively identify those at risk for asthma exacerbations5,6,7, making this a critical unmet need8,9,10. Metabolomics studies have been successful in developing clinical biomarkers11,12,13,14 and have the distinct advantage of reflecting both acute and long-term environmental influences in the context of underlying genetic predisposition. To date, asthma metabolomics studies have identified disruptions in several metabolic pathways, including sphingolipids, steroids, and microbial-derived metabolites15,16,17,18,19, that characterize the overall disease state across a spectrum of symptoms. Despite this progress, limitations in current asthma metabolomics studies persist, including small sample size, lack of validation, incomplete capture of metabolites20, and failure to consider the heterogeneity of the disease. Moreover, few studies have considered how metabolomics findings may be translated into clinical care. Metabolomics applications that characterize asthma heterogeneity and/or predict exacerbation risk may prove particularly useful for reducing the health burden and improving treatment efficacy. For metabolomics studies to effectively transition into clinical applications, it is imperative to move beyond discovery metabolomics and perform targeted analyses that delve deeply into the disrupted metabolic pathways and interrogate the clinical applicability of the initial findings. Targeted assays aid in identifying candidate biomarkers by providing robust data with a high level of confidence in metabolite identity and are more appropriate for adaptation into clinical applications in terms of cost, reproducibility and overall feasibility. Despite the logical progression to move beyond discovery metabolomics, targeted studies for multiple diseases, including asthma, remain limited.21,22 Among the targeted assays that did build on discovery metabolomics, there are notable successes in the development of clinical assays, for example, the application of ceramide ratios to assess risk for cardiac events23. Validation of the metabolic pathways implicated to date in asthma, followed by further characterization of their overall impact on the disease—in particular their potential to predict exacerbation risk—is a necessary next step to understand how these findings may have clinical impact. When discovery metabolomics implicates multiple pathways for a disease, identifying the interrelationships between these pathways may bring a more complete understanding of the underlying disease mechanisms for biomarker development24. Evaluating the impact of cross-pathway metabolite ratios for asthma is compelling for several reasons. These measures are straightforward to calculate and have been successfully used to develop predictive disease biomarkers previously25. Metabolite ratios have the advantage of providing biological insights into the interplay between two metabolic processes and are analytically more robust. The objective of this study was to build on the existing metabolomics literature to enable practical clinical translation in the context of real-life electronic medical records (EMRs). We utilized three well-characterized asthma cohorts, totaling 2513 participants, with up to 25 years of longitudinal data from EMRs and implemented discovery and independent replication across all stages of the study. We first validated the major dysregulated asthma pathways via discovery metabolomics. We then designed targeted assays to quantify the primary metabolites in the implicated pathways. We characterized the relationship between metabolites and multiple clinical traits for asthma and focused on the development and validation of a simple, cost-effective predictive biomarker panel to assess asthma exacerbation risk. In this process, we also used cross-pathway metabolite ratios to reflect the interrelationship of biological processes across metabolic pathways. This process established a framework that leverages global metabolomics followed by targeted assays to create a reliable and clinically relevant biomarker. This study included three asthma cohorts: the Mass General Brigham Biobank–Karolinska Asthma Study (MGBB-KAS, n = 1040), Mass General Brigham Biobank-Asthma (MGBB-Asthma, n = 610), and Mass General Brigham Biobank-Longitudinal Lung Function (MGBB-LLF, n = 823; formerly Omic Determinants of Longitudinal Lung Function in Asthma [ODOLLFA]), totaling 2513 individuals (Fig. Together, these cohorts represent a broad range of clinical and demographic characteristics common in adult asthma (Table 1.1, 1.2). The asthma cases in MGBB-KAS and MGBB-Asthma had moderate-to-severe disease, as indicated by elevated blood eosinophil and neutrophil counts, reduced pulmonary function, and higher IgE levels. Across both cohorts, patients exhibited varying patterns of inhaled corticosteroid (ICS) and oral corticosteroid (OCS) over the clinical course of disease, including the 5-year period after metabolomic profiling. MGBB-KAS The Mass General Brigham-Karolinska Asthma Study, MGBB-asthma The Mass General Brigham Biobank-Asthma, MGBB-LLF Mass General Brigham Biobank–Longitudinal Lung Function. This figure was created in BioRender. Discovery metabolomic profiling in the MGBB-KAS cohort identified 154 significant metabolite associations with asthma using a False Discovery Rate (FDR ≤ 0.05) correction for multiple testing. Of these metabolites, 46 replicated on a pathway level at nominal significance (p ≤ 0.05) in the MGBB-Asthma study, implicating eight metabolic pathways (Fig. Significant asthma-associated metabolites included microbial–derived metabolites, involved in four metabolic pathways that have been previously reported: 1) creatine metabolism26,27; 2) glycine, serine, and threonine metabolism28,29,30; 3) methionine, cysteine, taurine metabolism31,32; and 4) tryptophan metabolism33,34,35,36. We observed an inverse relationship between asthma status and steroid metabolites while adjusting for ICS treatment, which was supported by extensive prior literature16,31,37,38. Sphingolipids were strongly associated with asthma in the MGBB-KAS cohort, and their role in asthma has been substantiated in prior studies18,19,39; however, they were not included in the older global metabolomics platforms used for the MGBB-Asthma cohort, which prevented internal validation of these associations. We also observed increases in amino sugar and mannose metabolites with asthma diagnosis; however, these pathways were less substantiated in prior studies26,40. When considering these results in conjunction with the asthma metabolomics literature15,16,18, we identified steroids, sphingolipids, and microbial-derived metabolites for follow-up investigation via targeted assays. Points show adjusted odds ratios (ORs) from logistic regression; the error bars are 95% confidence intervals, and the center is the point estimate (OR). The x-axis is on a log scale. Two-sided Wald tests of β = 0 are reported with test statistic z = β/SE(β), and exact P values are reported in Supplemental Data 1. The significance levels of the regression analyses were FDR ≤ 0.05 for MGBB-KAS discovery and p value ≤ 0.05 for MGBB cohort replication. This figure summarizes the global profiling associations from the pathways we selected to study further via targeted assays. These pathways were selected based on their significant metabolite associations using MGBB-KAS in the global profiling, further supported by our validation study (MGBB-Asthma) and prior literature. We quantified 77 sphingolipids, 18 steroids, and 71 microbial-derived metabolites in serum samples from the MGBB-KAS cohort and observed 93 significant metabolite associations with asthma at an FDR ≤ 0.05 across six clinical measures related to asthma, including asthma diagnosis, pulmonary lung function, total serum IgE levels, and asthma exacerbations (Fig. The strongest associations across all targeted assays were observed for asthma exacerbations: 29 sphingolipids and 17 microbial-derived metabolites were positively associated with exacerbations, while 7 steroids were negatively associated with exacerbations. Overall, androgen, glucocorticoid, and progestogen metabolites were inversely associated with asthma diagnosis and exacerbations and positively associated with forced expiratory volume in one second (FEV1) and forced vital capacity (FVC). In particular, dehydroepiandrosterone sulfate (DHEAS), cortisone, and pregnenolone sulfate were associated with both asthma diagnosis (p value = 1.63 × 10−10, 5.33 × 10−9, 1.09 × 10−5, respectively) and decreased asthma exacerbations (p value = 3.11 × 10−8, 9.14 × 10−12, 7.22 × 10−5, respectively). While there were only a few significant microbial-derived metabolites associated with asthma (arginine: p value = 7.3 × 10−5; indol-3-propionate: p value = 3.4 × 10−6) and lung function (FVC: kynurenine, p value = 9.23 × 10−4; FEV1/FVC: betaine, p value = 2.87 × 10−4), 16 microbial-derived metabolites were associated with significant increases in exacerbations, including phenylacetylglycine, indoxyl sulfate, kynurenate, and quinolinate (p values = 6.66 × 10−10, 1.28 × 10−9, 8.67 × 10−9, and 3.69 × 10−8, respectively). Sphinganine-1-phosphate (sphinganine-1P) was positively associated with asthma (p value = 2.1 × 10−4). Many sphingolipid subclasses were associated with both increases in FEV1 and FVC; however, there were no associations with FEV1/FVC. Positive associations were observed with exacerbations across all sphingolipid classes, with Cer(d18:1/20:1), lactosylceramide(d18:1/14:0) (LacCer), and sphingomyelin(d18:1/14:0) (SM) exhibiting the most pronounced associations (p values = 2.48 × 10−6, 3.59 × 10−6, and 2.78 × 10−6, respectively). To further explore whether these associations were influenced by sex, we conducted sex-stratified analyses and present the results in Supplemental Data 3. We observed that males and females often exhibited distinct sets of significant metabolites for the same asthma phenotypes. In cases where overlapping metabolites were identified across sexes, the directions of association (β coefficients) were consistent, suggesting shared biological mechanisms despite potential sex-specific sensitivity. These findings indicate that while metabolite associations may vary by sex, the core directional relationships remain stable. A Targeted steroid panel; B Targeted microbial-derived metabolite panel; C Targeted sphingolipid panel; D Sphingolipid/microbial-derived metabolite ratios; E Sphingolipid/steroid ratios; F microbial-derived metabolite/steroid ratios. Points show adjusted regression coefficients (β) from linear regression; the error bars are 95% confidence intervals based on model-based standard errors, and the center is the point estimate. Two-sided t tests for β = 0 are reported with test statistic t(df) = value, and exact P values are reported in Supplemental Data 3. We observed a marked elevation in the percentage of significant associations between asthma traits and metabolite ratios compared to individual metabolites (Fig. We assessed the associations with asthma traits and 3898 sphingolipid to microbial-derived metabolite ratios (sphingolipid:microbial), 794 microbial-derived to steroid metabolite ratios (microbial:steroid), and 1248 sphingolipid to steroid ratios (sphingolipid:steroid). The overall association patterns identified consistent relationships between ratios of specific metabolite subclasses and asthma traits, either displaying strong overall significance or no significance at all (Fig. Thousands of associations between ratios and asthma traits were found to be FDR significant. The strongest associations among these were between asthma exacerbations and sphingolipid:steroid ratios, where 59.9% of these ratios tested were FDR significant. Ceramide/sphingomyelin to DHEAS ratios had the strongest associations with asthma exacerbations (p value = 1.63 × 10−26 – 1.63 × 10−22); however, significance with exacerbations was observed across a broad range of sphingolipid and steroid subclasses (Figs. The ratios of multiple sphingolipids to DHEAS were also associated with asthma diagnosis (p value = 3.86 × 10−16 – 8.56 × 10−14), while ratios between ceramides/sphingomyelins and estrone/deoxycorticosterone were negatively associated with log(IgE). These findings are biologically plausible given that sphingolipids, especially ceramides, are known to modulate steroidogenesis through the regulation of steroidogenic gene expression and intracellular signaling. Alterations in sphingolipid:steroid ratios may reflect imbalances in hormonal regulation and lipid signaling, both of which are critical in the inflammatory and immune processes underlying asthma. A The percentage of significant metabolite ratios by metabolite assay for different asthma clinical measures across the MGBB-KAS and MGBB-LLF. B Overview of the most significant metabolite ratios consistently identified across MGBB-KAS and MGBB-LLF. C Manhattan plots of associations between prevalent asthma exacerbation and the sphingolipid:steroid ratios across MGBB-KAS and MGBB-LLF. FEV1 forced expiratory volume in 1 second, FVC forced vital capacity, FEV1/FVC the ratio of FEV1 to FVC. Note: “Significant” refers to associations passing an FDR threshold of ≤0.05. Points = individual sphingolipid:steroid ratios (logistic regression). Two-sided t tests for β = 0 are reported with test statistic t(df) = value, and exact P values are reported in Supplemental Data 4. There were 873 FDR-significant associations between microbial:steroid ratios and asthma traits. The strongest microbial:steroid ratios were observed with asthma exacerbations (p values = 1.20 × 10−23 – 1.55 × 10−17), with DHEAS in the denominator, whereas a broader range of metabolite ratios were associated with asthma diagnosis (p values = 2.37 × 10−13 – 1.79 × 10−10), FEV1, FVC, and log(IgE) levels. There were 1603 FDR-significant associations between sphingolipid:microbial ratios and asthma traits. Specific metabolite subclasses were associated with asthma diagnosis (e.g., sphingosines, sphinganines, serotonin, indole-3-propionate), FEV1 (e.g., ceramides, sphingomyelins, kynurenine), FVC (e.g., sphingomyelins, quinolinate), and asthma exacerbations (e.g., ceramides, methylobutyrate, isobutyrate); however, there were no significant associations with FEV1/FVC or log(IgE). The significant metabolite ratio associations were independently replicated using the MGBB-LLF cohort (Supplemental Data 5). Similar association patterns were observed across both cohorts between clinical measures and pathway-level metabolite ratios. In both cohorts, the largest proportion of significant metabolite ratios was associated with asthma exacerbations, whereas a relatively small proportion of metabolite ratios was associated with lung function and log(IgE). The most significant metabolite ratios in MGBB-KAS replicated in MGBB-LLF, identifying ratios of specific metabolites and/or metabolite subclasses associated with distinct asthma traits: 1) Exacerbations: ceramides, sphingomyelin, DHEAS, isobutyrate; 2) FEV1 and FVC: ceramides, sphingomyelins, kynurenine, quinolinate, and testosterone; and 3) log(IgE): ceramides, deoxycorticosterone, indoles, and estrone. In both MGBB-KAS and MGBB-LLF, the highest proportion of FDR-significant associations was between 5-year incident and prevalent exacerbations and sphingolipid:steroid ratios. Manhattan plots of the exacerbations to sphingolipid:steroid ratios associations demonstrated that replication was consistent across the same subclasses of sphingolipid:steroid ratios in MGBB-LLF, with stronger overall statistical significance. Association p values in both cohorts ranged between 1.0 × 10−30–1.0 × 10−10 and were as low 1.39 × 10−36 in MGBB-LLF (Fig. Both cohorts showed the strongest associations for ratios with androgens, followed by glucocorticoids and progestogens. Significant ratios included multiple sphingolipid subclasses and were particularly elevated for those involving DHEAS, pregnenolone sulfate, androstenedione, cortisone, cortisol, and corticosterone in both cohorts. Further interrogation between asthma exacerbations and the sphingolipid:steroid ratios found that overall significance was driven by the combined impact of both the numerator and denominator of the metabolite ratios, rather than being driven by either the denominator or numerator alone (Supplemental Fig. Associations were not driven by outliers in the ratio distributions (Supplemental Fig. 4) and were more correlated with distinct clusters of the sphingolipid:steroid ratios and correlated with specific sphingolipid and/or steroid metabolites (Supplemental Fig. Results were robust to various covariate adjustments. To assess whether sphingolipid:steroid ratios can improve upon current clinical metrics in predicting exacerbations, we first evaluated the associations between clinical metrics available in the EMR and incident exacerbations. Despite nominal associations between higher neutrophil counts and exacerbators in the MGBB-KAS cohort (p = 0.028) and lower FVC among exacerbators in the MGBB-LLF cohort (p = 0.049), no clinical measures differed significantly between exacerbators and non-exacerbators across both cohorts (Fig. This suggests that current clinical characteristics alone are insufficient to distinguish which individuals will experience an asthma exacerbation in the next five years. We then assessed the value of including sphingolipid:steroid ratios to predict incident exacerbations alone and when used in combination with clinical measures. A The basic demographic and clinical characteristics between individuals with and without 5-year incident asthma exacerbations are presented for both the MGBB-KAS and MGBB-LLF cohorts; B ROC curves for incident asthma exacerbation classifiers across MGBB-KAS and MGBB-LLF; C Cox model results and cumulative incidence plots for the time until the first asthma exacerbation revealed by all 21 selected ratios in the predictive models in MGBB-KAS. FEV1 forced expiratory volume in 1 second, FVC forced vital capacity, FEV1/FVC the ratio of FEV1 to FVC, ROC receiver operating characteristic, AUC area under the curve, ICS inhaled corticosteroids. Curves show cumulative incidence = 1−Kaplan–Meier S(t) for strata Q1 and Q4. The center is the estimated cumulative incidence; shaded ribbons denote 95% pointwise confidence intervals. Using elastic net regression, the feature selection retained 21 sphingolipid:steroid ratios in the predictive model for incident exacerbations. Individual Cox models of all 21 sphingolipid:steroid ratios were significant and able to differentiate time until the first asthma exacerbation between exacerbation prone (Q1) and non-exacerbation prone (Q4) individuals in MGBB-KAS (Fig. Ratios with cortisone in the denominator, including Cer(d18:1/20:1):cortisone, HexCer(d18:1/24:1):cortisone, Sphinganine(d18:0):cortisone, and SM(d18:1/20:0):cortisone showed the strongest differentiations, with mean differences ranging from 265 to 366 days in the time until first exacerbation between the two groups (p value range = 2.50 × 10−12 – .02 × 10−9). When the metabolite ratios were available, we evaluated these associations in MGBB-LLF and identified that the biggest differences in the time until exacerbations were with ratios where DHEAS is in the denominator (p value range = 9.98 × 10−12 – 7.3 × 10−7), replicating these findings in MGBB-KAS. Ratios with cortisone did not replicate our findings in MGBB-KAS (Supplemental Fig. The prediction models that included the 21 sphingolipid:steroid ratios had the greatest predictive accuracy when compared with other clinical and baseline variables. The best predictive accuracy for 5-year incident asthma exacerbations, including 21 sphingolipid:steroid ratios serum IgE level, and baseline variables (race, ICS medication), achieving an area under the curve (AUC) of 0.901 (Supplemental Data 6). This predictive accuracy was validated in the MGBB-LLF cohort with a similar AUC of 0.893, despite the unavailability of 11-deoxycortisol in the MGBB-LLF cohort, which reduced the number of ratios to 16. In the predictive model, we included only individuals with complete data for selected input variables. Participants with missing values for candidate predictors (such as FEV1, IgE, or eosinophils) were excluded whenever these variables were part of model training (Table 2 summarizes missingness). When serum IgE was excluded from the model, the predictive model remained robust, with AUC values of 0.793 in MGBB-KAS and 0.724 in MGBB-LLF, indicating that the sphingolipid:steroid ratios alone have excellent predictive power. Recognizing the potential for selection bias when restricting the prediction modeling to asthma cases with specific asthma measures (e.g., IgE levels, FEV1, eosinophils, etc. ), we conducted additional analyses to evaluate how these patterns of missingness might impact our study conclusions. Specifically, we evaluated how other clinical characteristics of asthma cases may vary when IgE and/or FEV1 measures were missing. We found no significant differences in other asthma-related measures, such as eosinophil and neutrophil counts/percentage, lung function, between cases with and without these data. These findings suggest that missingness in IgE and/or FEV1 in the use of asthma cases is unlikely to introduce systematic bias, and therefore is unlikely to affect the development of the prediction model. In comparison, the baseline model of race and ICS medication achieved an AUC of 0.498. To further evaluate the predictive contribution of prior exacerbation history and traditional clinical markers, we constructed additional models incorporating OCS prescriptions, FEV1, eosinophils, and neutrophils. Prior exacerbation history (baseline model plus OCS prescriptions history) alone achieved an AUC of 0.726 in MGBB-KAS and 0.781 in MGBB-LLF. Including metabolite ratios alongside prior exacerbation history (Race, ICS, OCS, metabolite ratios) further improved the AUC to 0.845 and 0.844, respectively. Similarly, adding FEV1 to the race and ICS baseline model modestly improved prediction (AUC = 0.610–0.588), and combining FEV1 with metabolite ratios (Race, ICS, FEV1, metabolite ratios) achieved an AUC of 0.809 in MGBB-KAS and 0.733 in MGBB-LLF. In parallel, we constructed models incorporating eosinophils and neutrophils alongside metabolite ratios. These models resulted in lower AUC values (e.g., 0.795) compared to those that included IgE (AUC = 0.901), as shown in Supplementary Data 7. Models including only clinical markers (eosinophils, neutrophils, FEV1, or IgE without metabolite ratios) performed worse (AUC range: 0.515 to 0.654), all substantially lower than those incorporating metabolite ratios (Fig. These findings demonstrate that while blood eosinophils, FEV1, and prior exacerbation history are clinically informative, combining sphingolipid:steroid metabolite ratios with selected clinical variables yields the highest predictive accuracy for asthma exacerbations in our cohort. Using three large asthma cohorts with up to 25 years of EMRs, this study moves beyond global metabolomics profiling by following up primary findings with targeted assays to develop and validate an accurate predictive biomarker for 5-year incident asthma exacerbation. Throughout this process, we validated key metabolic pathways for asthma and characterized the relationship between metabolites and metabolite ratios with a spectrum of asthma traits. Our findings reinforce the important role of steroids, sphingolipids, and microbial-derived metabolites in asthma and identify reproducible relationships between distinct metabolite ratios and specific asthma traits. The strongest relationship was between sphingolipid:steroid ratios and exacerbations, where single ratios could differentiate high and low exacerbation proneness by up to one year. Based on these ratios, we constructed a validated predictive model for asthma exacerbations incorporating only 12 sphingolipids and 4 steroids, which outperformed conventional clinical predictors—including prior exacerbation history, FEV₁, eosinophil count, and IgE level—in both discovery and replication cohorts. The simple prediction models for exacerbations developed here offer cost-efficiency, increased accuracy, and straightforward implementation, suggesting a viable option for clinical assay development that may be useful in improving asthma treatment efficacy. After validating asthma-association metabolic pathways via global metabolomics, we selected steroids, sphingolipids, and microbial-derived metabolites for detailed investigation using targeted assays, ultimately focusing on the relationship between steroid and sphingolipid metabolism and 5-year asthma exacerbation risk. Sphingolipids and steroids have independently demonstrated important roles in asthma. Low ceramide levels in utero and early life have been linked to abnormal lung development and asthma41,42. While the specific mechanisms remain unclear, evidence suggests that environmental triggers, including allergens and infections, induce sphingolipid signaling and mast cell activation to instigate inflammation during asthma exacerbations43,44. Genetic studies have also shown that ORMDL3 polymorphisms, which inhibit serine palmitoyltransferase (SPT) and regulate sphingolipid homeostasis, contribute to asthma45. These findings demonstrate that sphingolipids play a complex, multifactorial role in asthma, yet more remains to be explored. The primary mechanisms of steroidogenesis are better understood and form the basis for corticosteroid treatment in asthma. Steroid deficiencies lead to increases in inflammation in the lung and are a risk factor for asthma exacerbations that start in utero and persist over the life course46,47,48. While ICS treatment effectively reduces lung inflammation and is efficacious for asthma control, evidence suggests that prolonged ICS treatment may suppress adrenal function, further exacerbating the underlying physiologic state16. The inherent link between endogenous steroid production and exogenous steroid treatment, therefore, complicates its role in the disease process. Nevertheless, this makes steroid metabolism an ideal candidate for optimizing asthma control. While we observed associations between asthma exacerbations and both sphingolipids and steroids, the associations and prediction accuracy were vastly improved with sphingolipid:steroid ratios. Importantly, the significant exacerbation associations replicated consistently across subclasses of sphingolipid:steroid ratios, with the most pronounced associations replicating for androgens, followed by glucocorticoid and progestogen species with that respective order across both cohorts. The use of metabolite ratios enables us to capture relative imbalances between metabolic pathways, offering insights that may not be apparent from individual metabolite levels alone. In asthma, where inflammatory lipid signaling and endocrine function are closely intertwined, sphingolipid-to-steroid ratios may reflect meaningful shifts in these biological processes. The strong associations between sphingolipid:steroid ratios and asthma exacerbations are further supported by the established interrelationship between these metabolite classes. Multiple sphingolipid species act as secondary modulators/regulators of steroidogenesis49,50,51, with distinct cellular functions that include regulating steroidogenic gene transcription52,53,54 and acting as both intracellular second messengers55,56 and extracellular paracrine/autocrine regulators57,58,59. Ceramides—produced through either de novo synthesis or via the hydrolysis of sphingomyelin - serve as the precursor to all sphingolipids and, therefore, play a particularly important role in sphingolipid metabolism. Ceramides are powerful in steroidogenesis because they can both directly and indirectly modulate steroid hormone production through their metabolism into other bioactive sphingolipids. Ceramides also act directly to suppress androgens and progestogens, further validating why these ratios may be clinically relevant60. Our findings broadly substantiate these mechanistic studies with particularly strong associations observed between ratios of ceramide, sphingomyelin, and Hex/Lac ceramide species to androgen, glucocorticoid, and progestogen species. Even single sphingolipid:steroid ratios demonstrated excellent discriminatory ability with direct clinical relevance, with single ratios effectively differentiating high and low exacerbators by as much as one year, highlighting the potential importance of this approach for clinical applications. While steroid regulation is important in asthma, the suppression of steroids is likely both a part of the disease etiology and a consequence of long-term ICS treatment. Yet, there is more to learn about how this pathway may impact asthma and its treatment. The relationship with the sphingolipids provides a new angle for how we might understand their impact. While individuals who have the same levels of specific steroids overall, when examined in combination with sphingolipids, some individuals are at much higher risk of an exacerbation, while others remain at low risk. This phenomenon sheds new light on steroids overall, where we can accurately discriminate between two people with similar cortisone or DHEAS levels into low and high-risk groups. This work suggests that the interaction between sphingolipids and steroids, rather than their isolated effects, may play a pivotal role in more fully understanding the condition and suggests that increased study into the interaction of these pathways is merited. While metabolomics research has primarily focused on the dysregulation of metabolites within a pathway, the interrelation between dysregulated pathways - captured in metabolite ratios—may be a crucial metric to describe disease-specific perturbations. Much of the progress with metabolite ratios to date has focused on alterations within a given metabolic pathway, while a relatively limited amount of epidemiological research has explored the relationship between metabolites across pathways. The use of within-pathway ratios, such as ceramide ratios for cardiovascular disease (CVD)61, kynurenine:tryptophan ratio for cancers62 and CVD63, and various lipid ratios for CVD64 haas marked some of the substantial advancements toward clinical translation. While between-pathway relationships are less well-studied, the approach we implemented here, by first limiting the metabolite ratios to metabolites from select pathways that are implicated in asthma, offers an initial approach to limiting the total number of ratios. Additional mechanistic work, such as what has been described, the interdependence between sphingolipids, steroids, and specific microbial-derived metabolites provides another approach to identify biologically relevant metabolite ratios65,66,67 that then may be studied further via targeted assays. However, the process of translating these findings into a clinical assay requires that the selected metabolites be vetted as viable candidates for clinical use. Sphingolipid:steroid ratios are not only predictive but would be relatively straightforward to implement into a clinical setting. These molecules are readily amenable to clinical assay development because they are abundant, stable, and the analytical methods for their quantification are relatively straightforward and inexpensive. If these ratios can accurately predict exacerbation risk within the next 6 months, that would enable the implementation of preventative measures to protect against an exacerbation. In addition, these ratios may identify those asthmatics who have poorly controlled asthma as demonstrated by increased numbers/frequency of exacerbation, suggesting that they should receive alternative treatment (e.g., biologics). In addition to the finding between spinngolipid:steroid ratios and exacerbations, the distinct patterns observed between specific metabolite ratios and other asthma traits offer several viable findings for follow-up. Specific ratios, such as ceramide/sphingomyelin to kynurenine/quinolinate, were strongly associated with lung function measures, but not other asthma traits. The phenotypic specificity between key metabolite ratios and specific clinical outcomes may also be important for understanding the biological mechanisms of asthma. Prior metabolomic studies have identified strong comorbidity across disease phenotypes68; however, there is a lack of specificity between metabolites and specific clinical outcomes. The specificity observed in the ratio associations suggests that these may serve as potential biomarkers for specific clinical outcomes. The current study has several limitations. First, the discovery metabolomics analyses conducted in the MGBB-KAS and MGBB-Asthma cohorts were performed using different platforms, resulting in discrepancies in metabolite annotations. Consequently, the validation of global association findings was performed at the sub-pathway level rather than at the level of individual metabolites. While targeted assays were applied in the MGBB-KAS cohort, the validation process in the MGBB-LLF cohort relied on global metabolomics data, which meant that not every metabolite or ratio identified in MGBB-KAS could be directly replicated in MGBB-LLF. This limitation underlines the challenges of cross-platform metabolomics analysis and highlights the importance of using targeted assays for validation, because they offer greater reproducibility and comparability across studies and laboratories69. Further follow-up work can focus on additional targeted assays covering other metabolic pathways. Second, while asthma clinical phenotypes used in this study were extracted from the EMR data, it is essential to recognize that real-world data, such as EMRs, might not match the quality of data derived from clinical studies or surveys. This discrepancy arises because EMRs are primarily tailored for clinical care rather than research. To address EMR-related concerns, we enhanced data reliability by calculating median laboratory test values over a five-year period around the data collection date, ensuring a robust dataset for our analysis. Our study utilizes EMR data from adult populations, which captures typical clinical practice. In our cohorts, IgE results were heterogeneously captured, leading to substantial missingness. Despite this, our analyses indicate that such missingness does not introduce meaningful bias into our predictive model. Moreover, the addition of IgE measurements significantly enhances the predictive accuracy of our model, suggesting IgE may reflect key physiological mechanisms underlying exacerbation susceptibility. Pending further validation, our results advocate for the regular assessment of IgE in adults with asthma to improve overall disease management. Additionally, while we used OCS prescriptions as a proxy for asthma exacerbations, we acknowledge that OCS may be prescribed for conditions unrelated to asthma. To assess the validity of this proxy, we conducted a correlation analysis between OCS prescription count and documented asthma exacerbation diagnoses in the EMR, which yielded a moderate-to-strong correlation (r = 0.573). This suggests reasonable specificity in our definition, although we recognize the need for future studies with more granular, temporally linked clinical data to further validate this endpoint. Furthermore, we observed variations in sample characteristics across the three cohorts, with the MGBB-KAS cohort utilizing serum for metabolomics measurements, in contrast to plasma used in the other cohorts. The strong validation suggested that our findings were robust to these cohort and matrix differences. Last, the study considered the impact of ICS treatment on steroid levels by adjusting the analytical model for ICS usage, addressing potential confounding effects and ensuring the validity of our results. Although faced with these constraints, we demonstrated robust replication of our findings. The use of targeted assays provided biochemical insights that were instrumental in refining disease prediction models and enhancing our understanding of the mechanisms underlying asthma and its diverse manifestations. These findings underscore the potential of metabolomics as a pivotal tool in the advancement of precision medicine. Moving forward, the next step will involve further refining these predictive models by incorporating additional asthma-related features, which will be crucial for enhancing the precision and application of metabolomics in precision medicine. As we move toward precision medicine, there is a need to translate the findings from large-scale omics studies into viable clinical biomarkers. This study focuses on bridging the gap to translation by focusing on asthma exacerbations, which comprise a major portion of the overall disease burden. We used the findings from global metabolomics to develop targeted assays that quantified select steroids, sphingolipids, and microbial-derived metabolites. We then developed a 5-year predictive model for asthma exacerbation risk using 21 sphingolipid-to-steroid ratios with high predictive accuracy that outperforms current clinical measures (discovery AUC = 0.90; replication AUC = 0.89). These findings underscore the value of metabolomics profiling and metabolite ratios to develop a practical, cost-effective clinical assay for asthma exacerbation risk that may improve patient care. Metabolite ratios have the potential to add to the predictive space and further inform us about biological mechanisms that contribute to important clinical outcomes. The overall goal of this study was to develop an accurate predictive model of exacerbations using global metabolomics profiling followed by selected targeted assays across three asthma cohorts (n = 2513) (Fig. The study design includes three stages: 1) Stage 1: Discovery metabolomics to identify asthma metabolic pathways related to asthma; 2) Stage II: Targeted assays for follow-up study and identification of top metabolites and metabolite ratios associated with asthma; 3) Stage III: Predictive model development for 5-year asthma exacerbation risk (Fig. Stage I employed two case-control studies, MGBB-KAS (n = 1080) and MGBB-Asthma, n = 610), with discovery circulatory metabolomics profiling to identify and validate top metabolite pathways associated with asthma. Combining these findings with existing metabolomics literature, top pathways were identified for further study via targeted assays. In Stage II, three targeted assays were developed to quantify 166 selected steroid, sphingolipid, and microbial-derived metabolites across the top pathways. Using asthma cases from two studies, MGBB-KAS and the MGBB-LLF (n = 823), the relationship between targeted metabolites and their ratios with asthma traits was evaluated to identify the strongest candidates for prediction modeling. In Stage III, a 5-year prediction model for asthma exacerbation risk was developed and compared with the performance of current clinical measures. The MGBB (https://biobank.partners.org) resulted from a project led by MGB (formerly known as Partners HealthCare) in which DNA, plasma, and serum samples were connected to clinical data from the EMR obtained from over 125,000 consented patients as of June 2021. Informed consent was obtained in written form from all participants. Patients involved in the biobank have provided consent through either in-person recruitment or electronic informed consent (eIC). Their involvement in the biobank allowed for blood sample data, EMR data, and survey data, including lifestyle, environmental, and family history information to be collected. The MGBB is linked with the MGB Research Patient Data Registry (RPDR), which stores EMR data on over 4.6 million patients in a SQL Server database, allowing researchers to query the RPDR through the MGBB Portal. The RPDR includes demographic data, diagnoses, procedures, medications, inpatient and outpatient encounter information, provider information, laboratory data, imaging and pathology data, and insurance information. EMR data are available from as early as 1999, and recruitment into the MGBB began in 2011. As such, individual participants may have up to 25 years of longitudinal EMR data, though the actual span varies by individual. For enrolled participants, extensive longitudinal EMR data, including relevant phenotypic information pertaining to asthma, can be extracted. MGB worked with the Institutional Review Board (IRB: 2014P001109) for approval for this collection of biospecimen data and the use of human participants for research through both the MGBB and the RPDR. The information obtained is approved for use in all types of research on human health, including genomics, biomarker analyses, epidemiology, and cell line creation. In this study, three independent asthma studies selected individuals from the MGBB using distinct study designs. Clinical phenotypes related to asthma, including lung function measures (FEV1, FVC and FEV1/FVC), blood differentials (eosinophil and neutrophil counts), IgE levels, ICS treatment, and asthma exacerbations defined via OCS treatments, were harmonized across all three cohorts. For some clinical phenotypes, such as lung function measures, blood differentials, and IgE levels, multiple values were often available per participant due to repeated clinical visits. To ensure temporal consistency with the metabolomics data, we selected the value closest in time to the serum sample collection date for each individual for analysis. The MGBB-KAS is a matched case-control study of asthma from the MGBB, consisting of 540 asthma cases and 540 non-asthmatic controls. An individual with asthma was defined by all the following criteria: 1) an asthma diagnosis defined by the asthma prediction algorithm70 in the RPDR (positive predictive value > 85%); 2) at least one oral steroid prescription or one inhaled steroid prescription (ICS of ICS/LABA ≥ 1). Control individuals were identified by the following: 1) no diagnosis of asthma (negative predictive value > 99%); 2) no ICS or OCS medications; 3) matched on age, sex, and self-reported race. All 1140 participants of MGBB-KAS were non-smokers and had serum samples available. Participants with missing serum collection dates and BMI information were excluded from subsequent analyses, along with their matched participants. As a result, 540 pairs of asthma cases and controls were included in the analyses (total n = 1080; Table 1). The MGBB-Asthma is an independent asthma case-control study in MGBB16 with 287 asthma cases and 323 control cases (total n = 610; Table 1), with no overlap from the MGBB-KAS cohort. A validated phenotyping algorithm70 in the RPDR was used for asthma diagnosis and identified 287 individuals with asthma (positive predictive value > 85%) and 323 controls (negative predictive value > 99%) to generate the MGBB-Asthma population (total n  =  610 individuals; Table 1). Non-fasting plasma samples for the MGBB-Asthma cohort were collected between October 2010 and March 2017 and were stored immediately (within 4 hours) in a −80 °C freezer. Controls were randomly selected from the pool of individuals without asthma with available plasma samples. The MGBB-LLF is also an independent MGBB cohort of 823 severe asthma cases, with no overlap from the MGBB-KAS and MGBB-Asthma cohorts. An individual with severe asthma was defined by all the following: 1) an asthma diagnosis defined by the asthma prediction algorithm70 based on the RPDR; 2) evidence of persistent disease activity, defined as at least three lung function assessments and/or ICS treatment in conjunction with adrenocorticotropic hormone (ACTH) testing. All participants of MGBB-LLF had available plasma samples and were non-smokers (total n = 823, Table 1). Samples were assayed on three liquid chromatography–mass spectrometry (LC-MS) platforms (HILIC-positive, lipidomics-positive, and lipidomics-negative) using an Agilent QToF 6550 interfaced with a Rapid Resolution separation module. For data quality control (QC) and pre-processing, coefficients of variation (CV%) were first computed, and features with CV% ≥ 25% were removed. Features missing ≥ 75% were also excluded. Missing values were imputed as half the minimum value across all samples for each feature. The resulting plots of principal component analysis (PCA) were examined, and the interquartile range (IQR) and skewness of each feature were computed. PCA was performed again, and the distribution of PCs according to demographic variables was examined. After QC, 2338 features from the HILIC-positive platform, 2672 features from the lipidomics-positive platform, and 1918 features from the lipidomics-negative platform were retained. Features with less than or equal to 30% missing were included in subsequent analyses (2317 from HILIC-positive, 2650 from lipidomics-positive, and 1885 from lipidomics-negative platforms). Metabolite annotations were assigned to features using the MS-DIAL software, resulting in 1902 metabolite identifications71,72. Unidentified features and xenobiotics were not included in further analysis. Untargeted global plasma metabolomics profiling was generated by Metabolon Inc. (Durham, North Carolina, USA). Batch variation was controlled for in the analysis. Sample preparation and global metabolomics profiling were performed according to methods described previously73,74,75. Metabolomic profiling was performed using four liquid chromatography tandem mass spectrometry (LC-MS) methods that measure complementary sets of metabolite classes described previously: 1) Amines and polar metabolites that ionize in the positive ion mode; 2) Central metabolites and polar metabolites that ionize in the negative ion mode; 3) Polar and non-polar lipids; 4) Free fatty acids, bile acids, and metabolites of intermediate polarity. All reagents and columns for this project were purchased in bulk from a single lot, and all instruments were calibrated daily for mass resolution and mass accuracy. Metabolite peaks were quantified using the AUC. Raw area counts for each metabolite in each sample were normalized to correct for variation resulting from instrument inter-day tuning differences by the median value for each run-day; therefore, the medians were set to 1.0 for each run. Metabolites were identified by automated comparison of the ion features in the experimental samples to a reference library of ~8,000 chemical standard entries that include retention time, molecular weight (m/z), preferred adducts, and in-source fragments as well as associated MS spectra and curated by visual inspection for QC using software developed at Metabolon, Inc.76. Identification of known chemical entities was based on comparison to metabolomic library entries of purified standards. Additional mass spectral entries were created for structurally unnamed biochemicals, which were identified by virtue of their recurrent nature. These compounds have the potential to be identified by future acquisition of a matching purified standard or by classical structural analysis. Missing metabolite values were imputed by replacement with half the minimum value for each metabolite in all samples. Metabolites with an IQRs of 0 were excluded from further analysis, with 904 metabolites remaining for the analysis. The remaining metabolites were subsequently log-10 transformed and Pareto-scaled. The sphingolipid profiles were quantified in all serum samples of MGBB-KAS participants (N = 1080) using LC-MS/MS. Details of the sphingolipid profiling method can be found in previous studies18,77. A total of 77 sphingolipids were reported, of which 42 were quantified using external calibration curves. For these 36 compounds, a pseudo-quantitation was performed based on compounds sharing similar structures and internal standards. Concentrations of sphingolipids were reported in nanomolar (nM). CV corresponding to each type of QC samples was calculated. Skewness of each sphingolipid was also calculated. The main sphingolipid subgroups included in this panel were dihydroceramides (dhCer), ceramides (Cer), ceramide-1-phosphates (Cer1P), sphingosines, sphingosine-1-phosphates (S1P), sphingomyelins (SM), hexosylceramides (HexCer), and lactosylceramides (LacCer), in addition to 3-ketosphinganine (3-KS), sphinganine, sphinganine-1-phosphate (Spa1P), and glucosylsphingosine. All sphingolipids had <7% missing values; thus, none were removed based on missing percentages. Sphingolipid concentrations were standardized (original value minus mean, then divided by standard deviation) to obtain effect estimates comparable among different sphingolipids. A separate batch of serum samples from the same participants was sent to Precion Inc. for quantification of a panel of microbial-derived metabolites that included 71 metabolites in the following pathways: short-chain fatty acids, tryptophan metabolism, dietary aromatic compounds, phenylalanine and tyrosine metabolism, polyamine metabolism, trimethylamine N-oxide biosynthesis, secondary bile acid metabolism, histidine metabolism, and vitamin biosynthesis. The microbial-derived metabolites panel analysis consists of two extractions of serum sample aliquots (50 and 100 μl). The extracts are analyzed with three different LC-MS/MS methods for the 71 analytes of interest. The concentration of microbial-derived metabolites was reported in μg/mL. This targeted panel was developed by Precion Inc., which described it as a “Microbial Metabolite Panel”. While some metabolites are indeed microbially produced, others are of dietary or host origin. To improve clarity, we use the term “microbial-derived metabolites” throughout the manuscript, and a comprehensive list of all 71 profiled metabolites is provided in Supplemental Data 2. Fifty-one metabolites missing≤30% were included in subsequent analyses. Microbial-derived metabolites' concentrations were standardized (original value minus mean, then divided by standard deviation) to obtain effect estimates comparable among different metabolites. A separate batch of serum samples from the same participants was sent to Precion Inc. for quantification of a steroid hormone panel, including 16 endogenous steroids in the following sub-pathways: androgens, estrogens, progestogens, glucocorticoids, and mineralocorticoids. The Comprehensive Steroid Hormone Panel utilizes 200 µl of serum or plasma, calibrated across analytes and conducted via three LC-MS/MS methods: Method 1 for aldosterone, estrone, and estradiol in ESI negative mode; Method 2 for non-conjugated steroids including prednisone and prednisolone in ESI positive mode; and Method 3 for steroid sulfates in ESI negative mode. Each method employs stable labeled internal standards for quantitation through linear regression analysis, with calibration and QC standards maintaining accuracy within ±15% of the nominal value and precision ≤10%. Steroid concentrations were standardized (original value minus mean, then divided by standard deviation) to obtain effect estimates comparable among different steroids. We summarized the basic demographic characteristics of participants in MGBB-KAS, MGBB-Asthma, and MGBB-LLF, stratified by asthma status (Table 1). Additionally, Table 2 provides a summary of the clinical measures relevant to asthma cases in MGBB-KAS and MGBB-LLF. We employed conditional logistic regression models using the clogit function from the survival R package to evaluate the associations between asthma status and individual metabolite features in the MGBB-KAS cohort, which followed a matched case-control framework. Matching was performed on age, race, and sex, and models were additionally adjusted for body mass index (BMI) values recorded closest to the serum collection date. To account for multiple testing, statistical significance was determined using a False Discovery Rate (FDR) correction. In the MGBB-Asthma cohort, we applied multivariable logistic regression models using the glm function (family = binomial) from the stats package, adjusting for age, sex, race, BMI, and smoking status, to replicate the significant asthma finding in MGBB-KAS. A finding was deemed to replicate if (1) the observed effect size (odds ratio, OR) aligned in direction with the initial association and (2) the association's p value was < 0.05. Using the global asthma metabolomics findings of the MGBB-KAS cohort and the scientific literature, we identified sphingolipid metabolism, steroid metabolism, and microbial-derived metabolites as priority pathways for further study. We expanded our study of these pathways through three targeted assays that quantified 77 sphingolipids, 71 microbial-derived metabolites, and 18 steroids. For these quantified metabolites, we used conditional logistic regression models implemented via the survival R package, accounting for the matched study design (matched on age, sex, and race). All participants were non-smokers, and BMI was included as an additional covariate to adjust for residual confounding. Additionally, we investigated the associations between targeted metabolites and asthma clinical measures among individuals with asthma. This investigation utilized multivariable linear (lm function, stats package) and Quasi-Poisson regression models (glm function with family = quasiposson) to evaluate various asthma clinical measures. Linear regression models focused on lung function measures such as FEV1, FVC, and the FEV1/FVC ratio, in addition to log(IgE). Quasi-Poisson models were employed to analyze the consumption of oral corticosteroids (OCS) over the past five years. Specifically, for lung function measures (FEV1, FVC, and FEV1/FVC ratio), we additionally adjusted for height. Given the smaller number and higher reliability of targeted metabolites, we applied a more stringent FDR ≤ 0.05 threshold to correct for multiple testing and ensure higher specificity of the results. In MGBB-KAS, the metabolite ratios were calculated using targeted metabolites from different panels, where the numerator and denominator originated from separate panels. The ratios were determined by dividing the raw concentration of one metabolite by another. Ratios were computed only when both the numerator and denominator metabolite values were available. Ratios with missing values of less than 30% were considered for further analyses. To ensure comparability and enhance model performance, all metabolite ratios were log₁₀-transformed prior to analysis. Sensitivity analyses demonstrated that this transformation improved model diagnostics, including residual normality and homoscedasticity, as shown in Supplemental Data 4. In this study, we focused on the ratio generated from three targeted panels (sphingolipids, microbial-derived metabolites, and steroids panels). The same multivariable linear/Quasi-Poisson regression models were utilized to assess the association between log-transformed metabolite ratios with asthma and its phenotypes within asthma cases, where the number of participants varied for each metabolite ratio. To correct for multiple testing, we applied a stringent FDR significance threshold (FDR ≤ 0.05). We extracted 62 sphingolipids, 46 steroids, and 118 microbial-derived metabolites to replicate our findings in the MGBB-LLF cohort. We calculated the ratios using raw values of these metabolites, ensuring the numerator and denominator were derived from the same three distinct categories as in MGBB-KAS. To ensure a normal distribution of the ratios, we applied a log10 transformation to all metabolite ratios. This process resulted in the generation of the same three varieties of ratios as identified in MGBB-KAS: Sphingolipid:microbial (number of ratios = 7316), sphingolipid:steroid (number of ratios = 2852), and microbial:steroid (number of ratios = 5428). To replicate the associations between metabolite ratios and asthma clinical measures, we employed regression models similar to those used in MGBB-KAS, adjusting for variables such as age, sex, race, and BMI (based on the closest record to the serum collection date), as well as ICS usage within the past five years. Additionally, we applied an FDR threshold of ≤ 0.05 in our statistical analysis. In our study, the presence of asthma exacerbations was assessed through the utilization of OCS medication. Participants who had not used OCS medication in the past five years were considered not to have experienced prevalent asthma exacerbations. Conversely, those whose OCS medication usage was equal to or exceeded the average (7 OCS prescriptions) within the last five years were classified as having prevalent asthma exacerbations. This same threshold was used to identify cases of incident asthma exacerbation over the next five years. The elastic net approach was employed to identify biomarkers indicative of asthma exacerbations in MGBB-KAS. In initial modeling, we included all FDR-significant metabolite ratios from both sphingolipid:steroid and microbial-derived metabolite:steroid classes to evaluate their combined predictive performance. After comparison, we subsequently focused on the sphingolipid:steroid ratios, which showed better performance and feasibility for assay development. Utilizing the glmnet R package, we trained binomial regression models for asthma exacerbations within the last 5 years on FDR-significant sphingolipid:steroid ratios alongside confounders such as age, sex, race, ethnicity, BMI (based on the most recent record before serum collection), and ICS usage. Through 10-fold cross-validation, an optimal lambda was selected based on the alpha parameter. Demographic variables like sex, race, and ethnicity were encoded using one-hot encoding. Ultimately, only those features with a non-zero coefficient were chosen as significant for identifying prevalent asthma exacerbation. This model was then tested to predict asthma exacerbations in the next 5 years in MGBB-KAS. To determine the discriminatory power of the single selected ratios for predicting incident asthma exacerbation, survival analysis was conducted using the Kaplan–Meier estimator to compare survival times between incident asthma exacerbation or not based on the single selected ratio through the survfit function from the survival R package. To further investigate the relationship between the single selected ratio and the time to first OCS medication, a Cox proportional hazards model was fitted using the coxph function. The model included covariates such as age at collection, BMI, gender, race, ethnicity, and ICS usage in the past five years. The mean time to the first OCS medication after the sample collection date was also calculated for participants with and without asthma exacerbation, based on the single selected ratio. All analyses are conducted in R version 4.1.278. Ethical approval was obtained from the IRB (2014P001109) of Brigham and Women's Hospital. Informed consent was obtained from all participants. We ensured the inclusion of participants irrespective of gender, race, or socioeconomic status. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Metabolomics data used in this study for the discovery cohort have been deposited in the Metabolomics Workbench (https://www.metabolomicsworkbench.org/data/DRCCMetadata.php?Mode=Project&ProjectID=PR002674) under Project ID PR002674. Requests for other data and materials will be reviewed by the cohort, and contact PIs Dr. Jessica Lasky-Su at rejas@channing.harvard.edu for the studies to determine if the request is subject to intellectual property or confidentiality obligations. We anticipate responding to requests within approximately 2–4 weeks. Data and materials that can be shared will be released using a Material Transfer Agreement. Appropriate IRB approvals may be required to access de-identified data, in particular, data from electronic medical health records. All custom R code used for data processing and statistical analyses is openly available at https://github.com/yuluc/metabolite_ratio_project. Bai, T. R., Vonk, J. M., Postma, D. S. & Boezen, H. M. Severe exacerbations predict excess lung function decline in asthma. Kelly, R. S. et al. Metabo-endotypes of asthma reveal differences in lung function: discovery and validation in two TOPMed cohorts. Boulet, L. P., FitzGerald, J. M. & Reddel, H. K. The revised 2014 GINA strategy report: opportunities for change. & Tang, D. D. Current understanding of asthma pathogenesis and biomarkers. Kaur, R. & Chupp, G. Phenotypes and endotypes of adult asthma: moving toward precision medicine. Carr, T. F. & Bleecker, E. Asthma heterogeneity and severity. Narasimhan, K. Difficult-to-treat and severe asthma: management strategies. & Liu, C. Machine learning for prediction of asthma exacerbations among asthmatic patients: a systematic review and meta-analysis. & Keshari, K. R. Metabolic analysis as a driver for discovery, diagnosis, and therapy. -j Mass spectrometry-based metabolomics in health and medical science: a systematic review. Lin, C. et al. Metabolomics for clinical biomarker discovery and therapeutic target identification. Kale, D., Fatangare, A., Phapale, P. & Sickmann, A. Blood-derived lipid and metabolite biomarkers in cardiovascular research from clinical studies: a recent update. Plasma metabolite profiles in children with current asthma. Kachroo, P. et al. Metabolomic profiling reveals extensive adrenal suppression due to inhaled corticosteroid therapy in asthma. Snowden, S., Dahlen, S. E. & Wheelock, C. E. Application of metabolomics approaches to the study of respiratory diseases. Chen, Y. et al. Sphingolipid classes and the interrelationship with pediatric asthma and asthma risk factors. Rago, D. et al. Characteristics and mechanisms of a sphingolipid-associated childhood asthma endotype. Julkunen, H. et al. Atlas of plasma NMR biomarkers for health and disease in 118,461 individuals from the UK Biobank. Villasenor, A., Eguiluz-Gracia, I., Moreira, A., Wheelock, C. E. & Escribese, M. M. Metabolomics in the identification of biomarkers of asthma. Vasile, V. C. et al. Ceramide scores predict cardiovascular risk in the community. Chen, Y. et al. Genomic atlas of the plasma metabolome prioritizes metabolites implicated in human diseases. Molnos, S. et al. Metabolite ratios as potential biomarkers for type 2 diabetes: a DIRECT study. Application of metabolomics in pediatric asthma: prediction, diagnosis and personalized treatment. Ho, W. E. et al. Metabolomics reveals altered metabolic pathways in experimental asthma. & Jang, A. S. N-acetylcysteine decreases airway inflammation and responsiveness in asthma by modulating claudin 18 expression. Song, J., Yao, L., Shi, J., Li, J. Aliyali, M., Poorhasan Amiri, A., Sharifpoor, A. & Zalli, F. Effects of N-acetylcysteine on asthma exacerbation. Comhair, S. A. et al. Metabolomic endotype of asthma. Fogarty, A., Broadfield, E., Lewis, S., Lawson, N. & Britton, J. Amino acids and asthma: a case-control study. Impaired tryptophan metabolism by type 2 inflammation in epithelium worsening asthma. Lee-Sarwar, K. A., Lasky-Su, J., Kelly, R. S., Litonjua, A. & Weiss, S. T. Gut microbial-derived metabolomics of asthma. Gostner, J. M., Becker, K., Kofler, H., Strasser, B. & Fuchs, D. Tryptophan metabolism in allergic disorders. Immune regulation and tryptophan metabolism in asthma. Huang, M. et al. Maternal metabolome in pregnancy and childhood asthma or recurrent wheeze in the vitamin D antenatal asthma reduction trial. Reinke, S. N. et al. Metabolomics analysis identifies different metabotypes of asthma severity. Levy, B. D. Sphingolipids and susceptibility to asthma. Zhou, Y. et al. Mannose receptor modulates macrophage polarization and allergic inflammation through miR-511-3p. & Obeid, L. M. Principles of bioactive lipid signalling: lessons from sphingolipids. Mendelson, C. R. Role of transcription factors in fetal lung development and surfactant protein gene expression. Maceyka, M. & Spiegel, S. Sphingolipid metabolites in inflammatory disease. Sturgill, J. L. Sphingolipids and their enigmatic role in asthma. Moffatt, M. F. et al. Genetic variants regulating ORMDL3 expression contribute to the risk of childhood asthma. How corticosteroids control inflammation: quintiles prize lecture 2005. & Lee, T. H. Glucocorticoid resistance in asthma is associated with elevated in vivo expression of the glucocorticoid receptor β-isoform. Ledeen, R. W. & Wu, G. Sphingolipids of the nucleus and their role in nuclear signaling. Ozbay, T., Merrill, A. H. Jr. & Sewer, M. B. ACTH regulates steroidogenic gene expression and cortisol biosynthesis in the human adrenal cortex via sphingolipid metabolism. Spiegel, S. & Milstien, S. Functions of the multifaceted family of sphingosine kinases and some close relatives. Budnik, L. T., Jahner, D. & Mukhopadhyay, A. K. Inhibitory effects of TNF alpha on mouse tumor Leydig cells: possible role of ceramide in the mechanism of action. Morales, V. et al. Intratesticular delivery of tumor necrosis factor-alpha and ceramide directly abrogates steroidogenic acute regulatory protein expression and Leydig cell steroidogenesis in adult rats. Ozbay, T., Rowan, A., Leon, A., Patel, P. & Sewer, M. B. Cyclic adenosine 5'-monophosphate-dependent sphingosine-1-phosphate biosynthesis induces human CYP17 gene transcription by activating cleavage of sterol regulatory element binding protein 1. Kwun, C. et al. Ceramide increases steroid hormone production in MA-10 Leydig cells. Meroni, S. B., Pellizzari, E. H., Canepa, D. F. & Cigorraga, S. B. Brizuela, L. et al. Sphingosine-1-phosphate stimulates aldosterone secretion through a mechanism involving the PI3K/PKB and MEK/ERK 1/2 pathways. Brizuela, L. et al. Sphingosine 1-phosphate: a novel stimulator of aldosterone secretion. Rabano, M. et al. Sphingosine-1-phosphate stimulates cortisol secretion. Interleukin-1 beta stimulates sphingomyelin hydrolysis in cultured granulosa cells: evidence for a regulatory role of ceramide on progesterone and prostaglandin biosynthesis. Hilvo, M., Vasile, V. C., Donato, L. J., Hurme, R. & Laaksonen, R. Ceramides and ceramide scores: clinical applications for cardiometabolic risk stratification. Increased serum kynurenine/tryptophan ratio correlates with disease progression in lung cancer. The kynurenine:tryptophan ratio as a predictor of incident type 2 diabetes mellitus in individuals with coronary artery disease. Barter, P. et al. HDL cholesterol, very low levels of LDL cholesterol, and cardiovascular events. The interplay between bioactive sphingolipids and steroid hormones. Ott, M. et al. Suppression of TDO-mediated tryptophan catabolism in glioblastoma cells by a steroid-responsive FKBP52-dependent pathway. Rasgon, N. et al. Neuroactive steroid-serotonergic interaction: responses to an intravenous L-tryptophan challenge in women with premenstrual syndrome. Langenberg, C., Hingorani, A. D. & Whitty, C. J. M. Biological and functional multimorbidity-from mechanisms to management. Torta, F. et al. Concordant inter-laboratory derived concentrations of ceramides in human plasma reference materials via authentic standards. Yu, S. et al. Toward high-throughput phenotyping: unbiased automated feature extraction and selection from knowledge sources. Naz, S. et al. Metabolomics analysis identifies sex-associated metabotypes of oxidative stress and the autotaxin-lysoPA axis in COPD. Meister, I. et al. High-precision automated workflow for urinary untargeted metabolomic epidemiology. Sha, W. et al. Metabolomic profiling can predict which humans will develop liver dysfunction when deprived of dietary choline. Evans, A. M., DeHaven, C. D., Barrett, T., Mitchell, M. & Milgram, E. Integrated, nontargeted ultrahigh performance liquid chromatography/electrospray ionization tandem mass spectrometry platform for the identification and relative quantification of the small-molecule complement of biological systems. Dehaven, C. D., Evans, A. M., Dai, H. & Lawton, K. A. Organization of GC/MS and LC/MS metabolomics data into chemical libraries. Ertl, P. Molecular structure input on the web. Akawi, N. et al. Fat-secreted ceramides regulate vascular redox state and influence outcomes in patients with cardiovascular disease. R: A Language and Environment for Statistical Computing. (R Foundation for Statistical Computing, 2021). is supported by R01HL155742 from the National Heart, Lung and Blood Institute, National Institutes of Health (NIH/NHLBI), USA. is supported by the NIH U01HG008685. Effort for J.A.L.S., J.H., and R.K. is supported by R01HL169300 from the NIH/NHLBI. is supported by U19AI168643 from the National Institute of Allergy and Infectious Diseases, NIH (NIH/NIAID). performed the QC and statistical downstream data analyses for the Mass General Brigham Biobank–Karolinska Asthma Study (MGBB-KAS) cohorts. performed the QC and statistical downstream data analyses for the validation cohort: Mass General Brigham Biobank–Longitudinal Lung Function (MGBB-LLF). (Ayobami Akenroye) contributed to ascertaining the inhaled and oral medications in MGBB-KAS and MGBB-LLF cohorts. contributed to phenotype curation, data cleaning, and quality assurance. All authors contributed to the review and critical revision of the final manuscript. These authors contributed equally: Yulu Chen, Pei Zhang. These authors jointly supervised this work: Craig E. Wheelock, Jessica A. Lasky-Su. Channing Division of Network Medicine, Department of Medicine, Brigham and Women's Hospital and Harvard Medical School, Boston, MA, USA Yulu Chen, Mengna Huang, Priyadarishini Kachroo, Qingwen Chen, Kevin Mendez, Meryl Stav, Nicole Prince, Sofina Begum, Andrea Aparicio, Tao Guo, Rinku Sharma, Su H. Chu, Rachel S. Kelly, Julian Hecker, Amber Dahlin, Scott T. Weiss, Michael McGeachie & Jessica A. Lasky-Su Unit of Integrative Metabolomics, Institute of Environmental Medicine, Karolinska Institute, Stockholm, Sweden Pei Zhang, Antonio Checa & Craig E. Wheelock Division of Allergy and Clinical Immunology, Department of Medicine, Brigham and Women's Hospital and Harvard Medical School, Boston, MA, USA Department of Respiratory Medicine and Allergy, Karolinska University Hospital, Stockholm, Sweden Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Correspondence to Craig E. Wheelock or Jessica A. Lasky-Su. is a scientific advisor to Precion Inc. and TruDiagnostic Inc. S.T.W. receives royalties from UpToDate and is on the Board of Histolix. Other authors have no relevant competing interests to disclose. Nature Communications thanks Sally Wenzel, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/. Chen, Y., Zhang, P., Huang, M. et al. Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Patient-facing large language models (LLMs) hold potential to streamline inefficient transitions from primary to specialist care. We developed the preassessment (PreA), an LLM chatbot co-designed with local stakeholders, to perform the general medical consultations for history-taking, preliminary diagnoses, and test ordering that would normally be performed by primary care providers and to generate referral reports for specialists. PreA was tested in a randomized controlled trial involving 111 specialists from 24 medical disciplines across two health centers, where 2,069 patients (1,141 women; 928 men) were randomly assigned to use PreA independently (PreA-only), use it with staff support (PreA-human), or not use it (No-PreA) before specialist consultation. The trial met its primary end points with the PreA-only group showing significantly reduced physician consultation duration (28.7% reduction; 3.14 ± 2.25 min) compared to the No-PreA group (4.41 ± 2.77 min; P < 0.001), alongside significant improvements in physician-perceived care coordination (mean scores 113.1% increase; 3.69 ± 0.90 versus 1.73 ± 0.95; P < 0.001) and patient-reported communication ease (mean scores 16.0% increase; 3.99 ± 0.62 versus 3.44 ± 0.97; P < 0.001). Co-designed PreA outperformed the same model with additional fine-tuning on local dialogues across clinical decision-making domains. Co-design with local stakeholders, compared to passive local data collecting, represents a more effective strategy for deploying LLMs to strengthen health systems and enhance patient-centered care in resource-limited settings. The growing burden of multimorbidity and aging populations has exposed vulnerabilities in healthcare delivery worldwide1,2,3. Health systems face increasing strain from fragmented infrastructure, under-resourced primary care and inefficient triage mechanisms4,5, challenges that are particularly acute in regions where self-referral practices bypass primary care for direct tertiary hospital access6,7,8,9,10. China's health system exemplifies this crisis: according to the 2023 Statistical Bulletin on China's Health Sector Development, hospital visits reached 4.26 billion in 2023 (11.5% annual increase), while only 59.2% of public hospitals offered appointment systems, driving inefficient care-seeking pathways that overwhelm outpatient services. This imposes a dual burden: specialists face patient consultations without referrals11, leading to prolonged diagnostic timelines12,13, compromised emotional support14,15 and elevated professional burnout16; concurrently, patients endure protracted waiting times and fragmented care17,18. Although interim solutions like nurse-led triage exist, they often lack the training for comprehensive patient assessment and chronic disease management19. Addressing these systemic inefficiencies in resource-limited settings requires scalable solutions that can transform strained clinical workflows. Large language models (LLMs) possess transformative potential to re-engineer hospital workflows and address the systemic inefficiencies amplified by escalating demand. However, current applications remain largely confined to support healthcare professionals in controlled settings, for example, responding to patient portal messages20,21, aiding clinical reasoning in experimental environments22,23 or improving medical directions in online pharmacies24, with limited integration into real-time clinical decision-making. Critically, evidence is lacking for LLM chatbots that directly interact with socioeconomically diverse patient populations while supporting both curative and caring aspects of medicine in high-volume clinical environments25. Bridging this gap requires overcoming two critical barriers: mitigating the systematic biases that arise when training patient-facing LLMs on local medical dialogues from resource-limited settings26,27, and establishing real-world evidence of their clinical utility within time-pressured hospital workflows28,29,30,31. While localized dialogues have enabled specialized applications, from patient-nurse interactions32 and mental health support33 to telemedicine service34, their direct use in resource-limited clinical environments risks replicating existing care deficits. Consequently, a shift toward simulated dialogues curated from standardized medical corpora is underway, moving beyond a reliance on raw local data23,35. Yet, the relative utility of co-design versus passive data collecting for meeting clinical needs remains unknown. This omission begs a central question: should LLMs reflect local practices or help reform them? The answer is critical for global health equity, as passively collected local dialogues may codify and even scale systemic inequities, from diagnostic shortcuts to sociocultural biases26,27. To bridge the gap between the potential of LLMs and their practical impact in resource-limited settings, we developed PreA (Pre-Assessment), an LLM chatbot (OpenAI; GPT-4.0 mini) for primary-to-specialist care transitions, using a multistakeholder participatory co-design approach36. We engaged diverse community and clinical stakeholders, including patients, care partners, community health workers, physicians, nurses and hospital administrators, to shape a tool that addresses real-world clinical and accessibility needs. The final PreA chatbot integrated a patient-facing chatbot with low-literacy accessibility features and a clinical interface that generates specialist referrals and supports evidence-based decision-making under time constraints (Extended Data Fig. The decision to deploy this co-designed version of PreA was empirically grounded, informed by a previous simulated experiment that directly compared the co-design approach with additionally fine-tuning the same model with local dialogues. We then evaluated the co-designed PreA in a multicenter, pragmatic, randomized controlled trial (RCT) to assess its effectiveness in facilitating primary-to-specialist care transitions. The trial was conducted across 24 medical disciplines at two academic tertiary medical centers in western China (The First Affiliated Hospital of Guilin Medical University and the Affiliated Hospital of Gansu Medical University). A total of 2,332 patients and their care partners were evaluated for eligibility, with 194 either opting out or being excluded for various reasons (Fig. This left 2,138 patients who were randomly assigned in 1:1:1 ratio to use PreA independently (PreA-only, n  = 712), use it with staff support (PreA-human, n = 713) or not use it (No-PreA, n  = 713). Subsequently, 69 patients opted out or were removed for various reasons. Flow diagram depicting the participant enrolment, intervention allocation, follow-up and data analysis. Most participants (1,620, 78.3%) were patients themselves, with the remainder being care partners. Demographically, fewer than half (881, 42.6%) were unemployed or retired, and 770 (37.2%) reported a monthly income below 2,000 RMB. Educational attainment among them was distributed as follows: below primary school (313, 15.1%), high school (1,073, 51.9%) and college or higher (683, 33.0%). These baseline covariates were well balanced across the three trial arms, with no significant differences in distribution (Table 1). Patients and their care partners in the PreA-only group spent approximately 3.51 ±1.50 min interacting with PreA, with no significant differences from those in the PreA-human group (3.48 ± 1.49 min; P = 0.72). They conducted, on average, no more than ten conversation turns, again with no significant differences between the two groups (PreA-only, 9.10 ± 1.37 versus PreA-human, 9.05 ± 1.26; P = 0.51). In their live clinical workflows, 111 specialist physicians reviewed PreA-generated and control (with age and sex only) referral reports immediately before patient consultations. These physicians spent an average of 0.25 ± 0.08 min reviewing PreA-generated reports (PreA-only and PreA-human), compared to 0.07 ± 0.06 min on control reports from the No-PreA group. We blindly assessed the effectiveness of the PreA consultation on outpatient workflows across three trial groups using data from the PreA platform and electronic hospital records. The PreA-only group had a significantly shorter consultation duration compared to the No-PreA group (PreA-only 3.14 ± 2.25 versus No-PreA 4.41 ± 2.77 min; P < 0.001; Fig. 2a), corresponding to a 28.7% (95% CI 22.7–34.8) relative reduction. No significant difference was observed between PreA-only and PreA-human groups (3.17 ± 2.87 min; P = 0.17). a, Histograms and box plots show the distribution consultation duration across the PreA-only (n = 691 participants), PreA-human (n = 689 participants), and No-PreA (n = 689 participants) groups. b, Box plots show patient throughput per shift for participating physicians and nonparticipating physicians based on 80 matched physician pairs. c, Radar plots show the patient-centeredness and care coordination metrics across the PreA-only (n = 691 participants), PreA-human (n = 689 participants) and No-PreA (n = 689 participants) groups, with five patient-reported metrics (ease of communication, physician attentiveness, interpersonal regard, patient satisfaction, and future acceptability) and one specialist-rated metric (care coordination). d, Bar charts show physician feedback at the end-of-shift questionnaires (n = 111 specialists). The left panel presents ratings of clinical decision support, workload reduction and facilitation of patient–physician communication. The right panel details the most valued features among physicians who rated its usefulness in decision-making as favorable or very favorable. The features include Interpretation (diagnostic report interpretation), Recording (efficient medical history elicitation and documentation), Across-discipline (simultaneous access to multiple specialties), Suggestion (preliminary diagnostic suggestions) and Communication (enhanced patient–physician communication skills). To evaluate the impact of PreA on the physician workload, we employed a matched-pairs analysis, comparing patients of participating physicians with those of nonparticipating physicians, matched on medical specialty, physician work shift timing and professional title. Patients of participating physicians cared for significantly more patients (28.54 ± 9.58) per shift compared to matched nonparticipating physicians (24.76 ± 9.42, P = 0.005; relative increase 15.3% (3.4–27.2); Fig. Given that participating physicians were exposed to PreA-only or PreA-human patients at a maximum frequency of two-thirds, this increase might represent a conservative estimate; on the other hand, physician could strategically control their workflow, thereby the actual impact of PreA on physician workload could be either more pronounced or less pronounced with universal PreA adoption. Despite caring for more patients, patients of participating physicians experienced similar waiting times compared to those of matched nonparticipating physicians (participating 33.54 ± 38.83 min versus nonparticipating 34.65 ± 36.92 min; P = 0.37). Outcomes here were self-reported by unmasked participants in the RCT (except for physicians masked between PreA-only and PreA-human arms) and measured using the prespecific survey questionnaire based on five-point Likert scales (Supplementary Tables 1 and 2). Patients and care partners in the PreA-only group reported significantly improved consultation experiences compared to the No-PreA group across the primary outcome, ease of communication: 3.99 ± 0.62 versus 3.44 ± 0.97; P < 0.001; relative increase 16.0%, 95% CI 13.5–18.5) and the four secondary outcomes: perceived physician attentiveness: 3.87 ± 0.85 versus 3.36 ± 1.04; P < 0.001; relative increase 15.1%, 95% CI 12.1–18.1), interpersonal regard (4.02 ± 0.73 versus 3.43 ± 1.05; P < 0.001; relative increase 17.2%, 95% CI 14.4–20.0), patient satisfaction (3.99 ± 0.69 versus 3.41 ± 0.98; P < 0.001; relative increase 17.0%, 95% CI 14.3–19.6) and future acceptability (3.79 ± 1.06 versus 2.81 ± 1.26; P < 0.001; relative increase 34.7%, 95% CI 30.4–39.1; Fig. No significant differences were found between the masked PreA-only and PreA-human groups across these dimensions (Extended Data Table 2). a, Agreement analysis between PreA reports and specialist notes. Cases were categorized by level of concordance: agreement (exact match, near-identical content or inclusion of accepted differentials), disagreement or blank (missing physician notes). b, Quality assessment of PreA reports versus specialist notes. The data represent the distribution of expert-evaluated quality across all available cases, including those with blank physician notes. For the primary outcome of referrals in facilitating specialist care, physicians reported a significantly higher value for PreA referral reports compared to the usual one (Care coordination: PreA-only 3.69 ± 0.90 versus No-PreA 1.73 ± 0.95; P < 0.001; Fig. 3b), corresponding to a 113.1% (95% CI 107.4–118.7) relative increase. No significant difference in perceived value was observed between the masked PreA-only and PreA-human groups (P = 0.45). At the end of each working shift, physicians provided feedback on the secondary outcomes, including the usefulness of PreA in their clinical decision-making (Fig. A majority reported PreA to be useful or very useful (64.9%, 72 of 111); among them, preliminary diagnostic suggestions (87.5%, 63 of 72) and efficient medical history acquisition (77.8%, 56 of 72) were identified as the most valuable features. Prespecified subgroup analyses, stratified by demographic and clinical characteristics, demonstrated consistent reductions in consultation duration. Notably, these reductions were observed across age groups, sex, educational attainment, work status, income levels, medical disciplines (medical medicine, surgery, mix of medical medicine and surgery, pediatrics), study sites (Guilin/Gansu) and participant type (patients/care partners), with PreA-only showing significant reductions compared to No-PreA, and no significant differences compared to PreA-human (Extended Data Figs. However, patient experience outcomes exhibited some variability across subgroups. While the PreA-only group generally reported superior consultation experiences compared to No-PreA, this effect was not uniformly observed. Specifically, high-income participants and those attending pediatric departments did not report significant differences in perceived physician attentiveness between the PreA-only and No-PreA groups (Extended Data Figs. Concerns regarding automation bias and anchoring, as in experimental contexts21, suggest clinicians may directly adopt LLM-generated assessments, potentially bypassing their clinical reasoning. To investigate this in our real-world trial, we examined whether physicians' clinical notes from the PreA-assisted groups exhibited distinct characteristics from those in the No-PreA group, as per the prespecified analysis. Classification analysis yielded near-random discriminability (F1 score 0.57; P = 0.81; ΔF1 < 0.02). This absence of systematic separability in the feature space of clinical notes provides compelling evidence against the direct adoption of LLM-generated content in this real-world clinical context. We further investigated this finding across five clinical domains of history-taking, physical examination, diagnosis, test ordering and treatment plans. Consistent with the overall findings, no significant difference was observed between the PreA-only and No-PreA groups (or PreA-only and PreA-human groups) in the five domains (P = 0.10–0.90; Extended Data Table 3). These findings collectively suggest that PreA-assisted medical consultation did not introduce detectable, systematic alterations in physician decision-making, either overall or within specific clinical domains. We performed a blind post hoc analysis comparing PreA referral reports to the subsequent physician clinical notes among the PreA-assisted groups. PreA-generated reports exhibited substantial agreement (exact match, near-identical content or inclusion of accepted differentials) with physician notes in 65.8% (95% CI 61.8–69.6) of history-taking, 66.7% (95% CI 62.7–70.4) of diagnoses, and 70.7% (95% CI 66.8–74.2) of test ordering recommendations (Fig. PreA reports show disagreement in only 2.8% to 5.7% of cases, while the remaining physician notes were absent that precluded direct comparison. Among cases exhibiting agreement or where physician notes were blank, PreA reports were rated significantly higher quality than physician notes in terms of completeness, appropriateness and clinical relevance across history-taking (PreA 4.73 ± 0.50 versus physician notes 2.93 ± 1.49), diagnosis (PreA 4.49 ± 0.82 versus physician notes 2.49 ± 1.25), and test ordering (PreA 4.55 ± 0.63 versus physician notes 3.28 ± 1.57; Fig. Intergroup analysis (PreA-only versus PreA-human) revealed no statistically significant differences in agreement rates and quality scores across all assessed domains. The choice of a co-designed chatbot for the RCT was informed by a previous simulated experiment that directly compared this approach against fine-tuning with local dialogues. For this experiment, we compiled a de-identified audio corpus of 515 patient–physician scenarios (199,145 Chinese words) collected across rural clinics and urban community health centers within the same 11 provinces as the co-design process. This dataset comprised general medical consultation interactions in geographically and socioeconomically diverse settings, with 51.7% (266 of 515) from rural areas and 77.9% (401 of 515) from the low-income regions. Mean consultation durations ranged from 1.55 to 3.98 min, and interaction lengths spanned 226.90 to 546.00 Chinese words per scenario (Fig. a, Geographic distribution and characteristics of the de-identified audio corpus comprising 515 patient–physician scenarios collected from the 11 provinces where local stakeholders participated in the co-design process. Provinces are categorized by income levels (high/low). Bar height represents the mean value, dots indicate individual data points, and error bars show 95% CIs. b, Quality score distributions for comparing co-designed PreA (n = 300 samples) with its local data-tuned counterpart (n = 300 samples), the same co-designed base model further fine-tuned on the primary care dialogues, across clinical evaluation domains. The co-designed model achieved significantly higher-quality rating scores than the data-tuned counterpart (the co-designed model further fine-tuned on the primary care dialogues) group across all domains: history-taking (without data-tuned 4.56 ± 0.65 versus data-tuned 3.86 ± 0.81; P < 0.001; Fig. Notably, the data-tuned model replicated systemic inefficiencies observed in real-world primary care, including omitting guideline-recommended history elements and demographic elements (for example, patient age and sex), and failing to provide appropriate tests and diagnoses (Supplementary Table 3). Mirroring real-world clinician patterns, the data-tuned model exhibited suboptimal adherence to diagnostic guidelines, failing to provide diagnoses (30.0%, 90 of 300) or suggest testing (39.3%, 118 of 300) when needed. Additionally, the data-tuned model mimicked an unfriendly tone similar to that of human clinicians. We developed and evaluated PreA, a co-designed LLM-based chatbot that streamlines primary-to-specialist care transitions by preparing patients for consultations and generating preconsultation referrals to specialists. In a pragmatic, multicenter RCT in China, PreA improved both operational efficiency and patient-centered care delivery in high-volume hospital settings compared to usual practice. The findings provide preliminary evidence for the clinical utility of co-designed LLMs within time-constrained clinical workflows, suggesting that co-design with local stakeholders is an effective strategy for deploying LLMs into clinical practice. The trial demonstrated that PreA enhanced both efficiency and patient-centeredness (a dual benefit rarely achieved in previous LLM deployments)20,21,22. Specialist physicians who received PreA-generated referral reports reduced their average consultation time by 28.7%, indicating that the tool enabled faster synthesis of clinical narratives and supported time-intensive decision-making. Indeed, the majority of specialists endorsed PreA's utility for rapid clinical synthesis, particularly valuing its preliminary diagnostic suggestions and medical history acquisition, which aligns with a recent qualitative investigation on physician views37. This efficiency gain, which could expand patient access or improve care quality, is particularly transformative in overloaded health systems where consultation lengths rank among the shortest worldwide38. Notably, this efficiency gain did not compromise (instead enhanced) both the cure-oriented and care-oriented medicine39,40, with physicians reporting improved care coordination and patients perceiving a more patient-centered experience. These efficiency cascades help to address core health system constraints identified in our co-design process, suggesting PreA's potential applicability in other health systems facing similar inefficiencies. The operational autonomy demonstrated by PreA, as evidenced by the equivalent performance of the PreA-only and PreA-human groups, carries important implications for scalability and cost-effectiveness for resource-constrained health systems41. Our matched-pair analysis revealed increased patient throughput per clinical shift even under partial PreA adoption, suggesting multiplicative system-level benefits when LLMs streamline preconsultation workflows. In resource-limited settings, such efficiency gains may substantially improve healthcare access, enhancing care equity. Additionally, the higher-quality scores of PreA reports position them as patient-specific templates that could alleviate the burden of clinical documentation. Future large-scale studies are needed to validate these potential benefits across diverse health systems. Our pre-trial ablation studies highlight an essential pathway toward equitable clinical AI: passively training LLMs on simply curated local dialogues risks perpetuating systemic care deficits, whereas participatory co-design could mitigate these risks and better align models with high-quality care objectives. The data-tuned model replication of suboptimal practices mirrors broader concerns that AI models trained on structurally biased clinical data exacerbate inequities in marginalized populations42,43. The co-designed PreA model, refined through input from local stakeholders, including patients, care partners, community health workers, primary care physicians and specialist physicians, outperformed the data-tuned model across all clinical domains. These findings underscore the architectural prioritization of local stakeholder agency through co-design over the passive assimilation of potentially biased natural dialogue data, advancing methodological approaches for equitable AI deployment in healthcare. Our study, alongside a concurrent trial demonstrating the efficacy of a co-designed chatbot for primary care in low-resource communities36, establishes participatory co-design as a versatile methodology for developing context-specific healthcare chatbots. While both RCTs employed similar co-design approaches, they target distinct clinical needs: while the primary care chatbot prioritized AI health literacy and accessibility for community home use, PreA was optimized for structured referral generation and time-efficient operation within high-volume specialist workflows. The resulting technical architectures and clinical applications consequently diverged, reflecting their distinct co-design processes and stakeholder priorities. These complementary findings demonstrate how co-design principles can be adapted to develop tailored LLM solutions for diverse healthcare contexts, serving various patient populations and clinical objectives. In contrast to previous research that has often framed LLMs as physician-interaction diagnostic entities22,27,44,45,46, our findings show that a co-design approach, involving the iterative alignment of LLMs with the prioritized needs of local stakeholders, represents the necessary next step, moving beyond technical promise to clinically integrated, equity-focused AI tools28. The streamlined integration of PreA's outputs with specialist cognitive workflows resulted in significant reductions in consultation time and enhanced patient experience across demographic and socioeconomic strata. Critically, these findings challenge the prevailing narrative that medical AI tools inherently depersonalize medicine47, instead positing that co-designed LLM deployments could empower clinicians to prioritize patient-centered care when freed from cognitive burdens. Furthermore, while previous LLM models may have achieved success within narrow, siloed domains22, PreA's demonstrated cross-disciplinary effectiveness, spanning both surgical and medical specialties, underscores its potential to unify currently fragmented care pathways across medical disciplines28. Several limitations warrant consideration when interpreting our findings. The generalizability of our time-reduction findings may be context-dependent, as our study was conducted in high-volume, resource-limited hospital settings. The effectiveness of PreA is intrinsically tied to this environment of high clinical demand and standardized workflows, and validation in diverse healthcare systems is warranted. Furthermore, the single-blinded, pragmatic trial design, while reflecting real-world conditions where patients would naturally know their preconsultation experience, introduces potential performance bias as patients were aware of their group assignment; however, several factors mitigate this concern: the concordance of findings across objective and subjective outcome assessments, the absence of significant differences in clinical documentation across trial arms and the alignment of control group consultation times with established practice patterns. Although co-design demonstrated advantages over local data fine-tuning for mitigating biases in LLM development, this approach remains constrained by data quality limitations in health resource-limited settings. Future comparative studies should evaluate co-design versus emerging high-quality primary care dialogue datasets to better understand their relative strengths and applications. The systemic documentation gaps48, evidenced by missing physician notes, represent both a limitation and an important finding. While our analytical methods account for this missingness, future implementations could leverage PreA reports as documentation aids to address this widespread challenge in high-volume settings. Moreover, while PreA demonstrated potential as a primary-to-specialist care transition aid, its transition into home-based use would represent an optimal future direction that requires addressing systemic barriers, including AI health literacy, connectivity limitations and cross-institutional data sharing, as indicated by other work on co-designing LLMs for primary care settings36. This study provides preliminary evidence for integrating patient-facing LLMs into hospital workflows. While larger multicenter trials with longer follow-up are needed to establish sustained benefits, cost-effectiveness and generalizability, our findings mark a significant step forward. The demonstrated improvements in workflow efficiency and patient–physician experience indicate that co-designed chatbots can reallocate clinician effort from routine data processing toward more nuanced and meaningful patient interactions. This work underscores co-design with local stakeholders as an effective strategy for deploying LLMs to strengthen health systems and enhance patient-centered care in resource-limited settings. The Chinese Academy of Medical Sciences and Peking Union Medical College and the local medical ethics committee of the First Affiliated Hospital of Guilin Medical University approved the study. We obtained informed consent from all participants in this study. All participants were informed that this was an exploratory experiment, and the results should not be interpreted as direct guidance for clinical interventions at this stage. This study implemented stringent data protection measures, ensuring that all data were anonymized and encrypted to protect privacy. PreA's architecture, derived from the co-design with local stakeholders, integrates a patient-facing chatbot and a clinician interface (Extended Data Fig. The patient interface collects medical history via voice or text, while the clinical interface generates structured referral reports. Within the consultation workflow, patients or their designated caregivers interacted with PreA first; it then generated a referral (Supplementary Table 4) for specialists to review before their standard consultation. Co-design workshops revealed that standard clinical documentation for common conditions often lacks the granularity for personalized care and omits critical patient-specific details like pre-existing comorbidities. As such, PreA referral reports were intentionally designed to bridge this gap by synthesizing comprehensive, patient-specific information to facilitate rapid documentation and diagnostic decision-making. The architecture was also engineered to support both multidisciplinary consultation (prioritized by primary care physicians) and evidence-based diagnostic reasoning (emphasized by specialist physicians). PreA's consultation logic underwent a two-cycle co-refinement process to achieve broader utility across diverse socioeconomic patient populations and adherence to World Health Organization (WHO) guidelines for equitable AI deployment (Extended Data Fig. The first cycle involved adversarial testing with 120 patients and caregivers, 36 community health workers, 15 physicians and 38 nurses from urban and rural areas across 11 provinces (Beijing, Chongqing, Gansu, Hubei, Shaanxi, Shandong, Shanxi, Sichuan, Guangxi, Inner Mongolia and Xinjiang). This participatory refinement enhanced real-world contextualization and mitigated potential disparities in health literacy and workflow integration49. The second cycle employed a virtual patient simulation, specifically modeling low-health-literacy interactions to further optimize the model against co-designed evaluation metrics. Subsequent sections provide further methodological details. The patient-facing chatbot employs a two-stage clinical reasoning model: inquiry and conclusion. During the inquiry stage, the model was trained to conduct active, multiturn dialogues to gather comprehensive health-related information, adhering to standard guidelines on general medical consultation. In the conclusion stage, the model generated 1–3 differential relevant diagnostic possibilities, each with supporting and refuting evidence to enhance diagnostic transparency and mitigate cognitive anchoring risks50,51. PreA was configured to generate a referral report for primary-to-care transitions. The report included patient demographics, medical history, chief complaints, symptoms, family history, suggested investigations, preliminary diagnoses, treatment recommendations and a brief summary aligned with clinical reasoning documentation assessment tools (Supplementary Table 4). To ensure accessibility, the platform supports shared access for patients and their caregivers52. An LLM-driven agent performs real-time intention analysis to facilitate empathetic communication and simplify language for low-literacy users, with outputs formatted as JSON for streamlined processing. To improve clinical utility under time constraints, the model was optimized to balance comprehensive data gathering with clinical time constraints, targeting 8–10 conversational turns based on local stakeholder feedback. Primary care physician input drove the incorporation of high-yield inquiry strategies, which in pilot testing reduced consultation times by approximately half (within 4 min). In the adversarial stakeholder testing cycle, we employed prompt augmentation and agent techniques to refine the model, aligning the chatbot with WHO guidelines for ethical AI in primary care while preserving clinical validity53. An evaluation panel consisting of community and clinical stakeholders and one AI-ethics-trained graduate student, conducted iterative feedback cycles, focusing on mitigating harmful, biased or noncompliant outputs via adversarial testing. In the simulation-based refinement cycle, we used bidirectional exchanges between PreA and a synthetic patient agent to enhance consultation quality. We synthesized 600 virtual patient profiles using LLMs grounded in real-world cases; 50% (n = 300) required interdisciplinary consultation to reflect complex care needs. Five board-certified clinicians validated all profiles for medical plausibility and completeness (achieving 5 of 5 consensus). The patient agent was built on a knowledge graph architecture54, formalizing patient attributes (demographics, medical history and disease states) as interconnected nodes. The agent was further instructed to emulate common consultation challenges identified by community stakeholders in the first cycle. Interactions concluded automatically upon patient acknowledgment or after ten unresolved inquiry cycles. We randomly chose 300 profiles for refinement and reserved the remainder for comparative simulation studies. The co-design process identified five consultation quality domains for refinement: efficiency (meeting the patient's demanding time lengths), needs identification (accurate recognition of patient concerns), clarity (concise and clear inquiries and responses), comprehensiveness (thoroughness of information) and friendliness (a respectful and empathetic tone). PreA's performance was rated across these metrics by a panel of five experts (two primary care physicians, two specialists (one in internal medicine and the other in surgical medicine) and one AI-ethics-trained graduate student). Separately, two primary care physicians assessed referral reports for completeness, appropriateness and clinical relevance using a co-designed, five-point Likert scale (Supplementary Table 5). Scores below 3 triggered further iterative refinement. We collected audio recordings of primary care consultations from rural clinics and urban community health centers across the 11 Chinese provinces. Provinces were categorized as low-income and high-income based on whether per capita disposable income was below or above the national average (National Bureau of Statistics of China). Local co-design team members who live in these areas manually calibrated the transcripts to ensure validity, as the raw data contained noisy, ambiguous language, interruptions, ungrammatical utterances, nonclinical discourse and implicit references to physical examinations. All conversational data collected was rigorously de-identified in compliance with relevant regulatory standards (HIPAA) before data analysis. We conducted a comparative simulation study to evaluate the incremental utility of integrating these localized dialogues. Two model variants were compared: the co-designed PreA model and a local data-tuned counterpart, created by fine-tuning the PreA model (OpenAI; ChatGPT-4.0 mini) on the processed primary care dialogues. Consequently, when behavioral cues conflicted, the model would preferentially adhere to patterns learned from the fine-tuning data. The virtual patient experiment utilized 300 unused patient profiles to evaluate clinical decision impacts (history-taking, diagnosis and test ordering). Referral reports from both variants were blindly evaluated by the same expert panels as in the PreA development, using validated five-point Likert scales for completeness, appropriateness, and clinical relevance (Supplementary Table 5). Inter-rater reliability for these assessments was high (κ > 0.80), and group comparisons were conducted using the two-tailed nonparametric Mann–Whitney U-tests. In this pragmatic, multicenter RCT, patients were randomized to use PreA independently (PreA-only), with staff support (PreA-human) or not use it (No-PreA) before specialist consultation. The PreA-human arm was included to assess PreA's autonomous capacity. The primary comparison was between the PreA-only and No-PreA arms, with a secondary comparison between PreA-only and PreA-human arms. The trial's primary end points were to evaluate the effectiveness of the PreA in enhancing operational efficiency and patient-centered care delivery in high-volume hospital settings, as measured by consultation duration, care coordination, and ease of communication. Examples of patient interaction with PreA are provided in Supplementary Tables 6–8. Participants must demonstrate a need for health consultation or express a willingness to engage in PreA health consultations. Other inclusion criteria were (1) aged between 20 years and 80 years; (2) visit the participating physicians at the study medical centers; (3) eligible for communicative interaction via mobile phone; (4) eligible to complete the post-consultation questionnaires; and (5) have signed informed consent. Exclusion criteria were (1) the presence of psychological disorders; (2) any other medical events that are determined ineligible for LLM-based conversation; and (3) refusal to sign informed consent. No co-design stakeholders participated in the RCT. In the PreA-human group, participants were informed that PreA's interface was similar to WeChat, which has been used by hospitals for patient portal registries and hospital visit payments, and were offered technical support. For both PreA-assisted arms, a PreA-generated referral report was provided to specialist physicians for review via the patient's mobile phone before any face-to-face interaction. This design was implemented to prevent direct copying of content into clinical notes. Physicians were requested to rate the report's value for facilitating care. In the No-PreA control arm, physicians reviewed routine reports containing only patient sex and age. Following consultations, patient and care partner experiences were captured via a post-consultation questionnaire (Supplementary Table 1). Physicians provided feedback at the end of their shifts (Supplementary Table 2). The primary outcomes were consultation duration, physician-rated care coordination, and patient-rated ease of communication. These metrics were selected based on co-design feedback, which identified time efficiency and care coordination as critical for adoption in high-workload settings, and are established proxies for clinical effectiveness and patient-centered care38. Secondary outcomes included physician workload (measured as patients seen per shift and compared between participating and matched nonparticipating physicians); patient-reported experiences of physician attentiveness, satisfaction, interpersonal regard and future acceptability; physician-reported assessments of PreA's utility, ease of communication, and workload relief; and clinical decision-making patterns, derived from a quantitative analysis of clinical notes. Consultation duration, patient volume and clinical notes were extracted from the PreA platform and hospital electronic records. Physician-perceived care coordination was measured via a five-point Likert scale rating the helpfulness of the PreA report in facilitating care. Patient-reported and other physician-reported outcomes were collected using prespecified, five-point Likert scale questionnaires (Supplementary Tables 1 and 2). These instruments demonstrated robust internal consistency (Cronbach's α > 0.80 for all domains) and face validity, established through iterative feedback from 20 laypersons and five clinicians to ensure relevance to outpatient contexts. Clinical notes were extracted from all three trial arms and included five core clinical reasoning domains: history-taking (chief complaint, history of present illness and past medical history), physical examination, diagnosis, test ordering and treatment plans. The target minimum sample size of 2,010 participants (670 per study arm) was prespecified based on a power analysis using preliminary data from the pilot study of 90 patients. This minimum target sample size ensured sufficient power (>80%) for the primary outcome at a significance level of 0.05. The clinical research team approached adult patients from waiting rooms who were scheduled to see the participating physicians. For pediatric patients or adults without a mobile device, their caregivers were contacted. Interested individuals received a comprehensive study description, which emphasized the exploratory nature of the research and clarified that any advice rendered by PreA serves solely as a reference and should not be utilized as a definitive basis for disease therapy. After providing informed consent and having their questions addressed, eligible individuals who met the inclusion and exclusion criteria were formally enrolled. Recruitment was conducted from 8 Feb to 30 April 2025. Participants were allocated to one of the three groups using an individual-level, computer-generated randomization sequence without stratification. Allocation was concealed to prevent selection bias. This trial was single-blinded: while the patients knew their group assignments (PreA-only, PreA-human, or No-PreA), the physicians were uninformed about the PreA-intervention groups (PreA-only or PreA-human). Furthermore, research staff involved in data analysis remained blinded to group assignments throughout the study. We assessed baseline covariate balance across the three groups using an ANOVA for continuous variables and a chi-squared test for categorical variables. For intergroup comparisons, we evaluated the distribution of scale values; two-sample Student's t-tests with unequal variances were used for approximately normal data, while a nonparametric Mann–Whitney U-test was applied to skewed distributions. All tests were two-tailed with a significance threshold of P < 0.05. The Benjamini–Hochberg procedure was applied to correct for multiple comparisons. Secondary analyses evaluated the consistency of findings across demographic and socioeconomic subgroups. Python v.3.7 and R v.4.3.0 were used to perform the statistical analyses and present the results. We conducted a matched-pairs analysis to assess the impact of PreA on physician workload. A second matched-pairs analysis was conducted to assess the effect on patient waiting times. For this analysis, we matched participating physicians to a distinct control group on the same covariates (replacing working shift with working week to accommodate the matching on patient volume) and the number of patients seen per shift. For both analyses, statistical significance between participating physicians and their matched controls was assessed using two-sided Wilcoxon signed-rank tests to account for matched data. We performed a prespecified classification analysis to detect systematic differences between clinical notes from PreA-assisted arms (PreA-only and PreA-human groups) and the No-PreA arm. Notes were partitioned into training and test sets (2:1 ratio). A binary classifier was trained to distinguish PreA-assisted from No-PreA notes, with classification performance evaluated using the F1 score, defined as the harmonic mean of precision and recall F1 Score=TP/(2TP + FP + FN), where TP, FP and FN denote true positives, false positives and false negatives, respectively. Under the null hypothesis of no intergroup differences, classifier performance would be random. A statistically significant F1 score exceeding this baseline (ΔF1 > 0.02) would indicate distinguishable clinical decision-making patterns attributable to PreA-assisted notes. The classifier was trained in two stages. First, a Med-BERT encoder generated contextualized embeddings of the clinical notes55. Second, a binary classification layer was trained on these embeddings using supervised SimCSE56, a contrastive learning approach that minimized embedding distance within the PreA-assisted group while maximizing the distance to the No-PreA group. Statistical significance was assessed with one-sided bootstrap tests (1,000 samples). A prespecified domain-specific analysis further compared clinical decision-making across five domains. For structured non-normal count data (number of diagnoses, number of tests ordered and number of treatments), documented counts were compared between groups using nonparametric Mann–Whitney U-tests. A post hoc analysis evaluated the concordance and quality of PreA-generated referral reports versus physician-authored clinical notes. Agreement was defined as substantial alignment (exact match, near-identical content or inclusion of accepted differential diagnoses). The same expert panel from the comparative simulation studies (two board-certified primary care physicians and two senior residents) performed blinded ratings of report and note quality on a five-point Likert scale for completeness, appropriateness and clinical relevance (Supplementary Table 3). They assessed three domains relevant to primary care referrals: history-taking, diagnosis and test ordering; physical examination and treatment plans were excluded. Each case was evaluated by two experts. Group comparisons were performed using a nonparametric Mann–Whitney U-test. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Source data are provided in Tables and Extended Data Tables and can be accessed via the code repository (https://github.com/ShashaHan-collab/PreA-OutpatientRCT)57. Raw conversation data are not publicly available due to the need to protect participant privacy, in accordance with the ethical approval for this study. Anonymized, nondialogue individual-level data underlying the results can be requested by qualified researchers for academic use. Requests should include a research proposal, statistical analysis plan and justification for data use, and can be submitted via email to S.H. All requests will be reviewed by the Chinese Academy of Medical Sciences & Peking Union Medical College and the ethics committee of the First Affiliated Hospital of Guilin Medical University. Review of the proposals may take up to 2 months, and approved requests will be granted access via a secure platform after execution of a data access agreement. Comparative statistical analyses were detailed in the paper. Code for classification analysis and data visualization can be found at https://github.com/ShashaHan-collab/PreA-OutpatientRCT (ref. The PreA chatbot is not publicly available as it is the subject of ongoing commercial licensing discussions and is protected intellectual property held by the Chinese Academy of Medical Sciences & Peking Union Medical College, intended for development as a regulated medical device. To preserve commercial viability and prevent the unregulated use of a patient-facing clinical tool, public release is not permitted at this time. To support validation and collaborative academic research, the core PreA model can be made available to qualified researchers upon a formal request to S.H. (hanshasha@pumc.edu.cn), subject to a data-sharing agreement, ethical approvals and a commitment to appropriate safety protocols. Multimorbidity: A Priority for Global Health Research (Academy of Medical Sciences, 2018). Michel, J. P., Leonardi, M., Martin, M. & Prina, M. WHO's report for the decade of healthy ageing 2021–30 sets the stage for globally comparable data on healthy ageing. Han, S. et al. Multimorbidity progression and the heterogeneous impact of healthy ageing risk factors: a multicohort study. Primary care: an increasingly important contributor to effectiveness, equity, and efficiency of health services. High-quality health systems in the Sustainable Development Goals era: time for a revolution. Patient self-referral patterns in a developing country: characteristics, prevalence, and predictors. Nakayuki, M., Basaza, A. H. D. & Namatovu, H. K. Challenges affecting health referral systems in low-and middle-income countries: a systematic literature reviews. & Njuguna, S. Challenges facing implementation of referral system for quality health care services in Kiambu county, Kenya. Godlee, F. Put patients first and give the money back. Gordon, G. H., Baker, L. & Levinson, W. Physician-patient communication in managed care. Effective physician-patient communication and health outcomes: a review. Haskard K. B., Williams S. L. & Dimatteo M. R. Physician-patient communication: psychosocial care, emotional well-being, and health outcomes. In Communicating to Manage Health and Illness (eds Brashers, D. E. & Goldsmith, D.) 23–48 (Routledge, 2009). De Hert, S. Burnout in healthcare workers: prevalence, impact and preventative strategies. & Lu, G. Factors associated with outpatient satisfaction in tertiary hospitals in China: a systematic review. Elliott, D. J., Young, R. S., Brice, J., Aguiar, R. & Kolm, P. Effect of hospitalist workload on the quality and efficiency of care. The roles and responsibilities of general practice nurses in China: a qualitative study. Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. Goh, E. et al. GPT-4 assistance for improvement of physician performance on patient care tasks: a randomized controlled trial. Pais, C. et al. Large language models for preventing medication direction errors in online pharmacies. Artificial intelligence, ChatGPT, and other large language models for social determinants of health: Current state and future directions. The future landscape of large langÿuage models in medicine. High-performance medicine: the convergence of human and artificial intelligence. Testing and evaluation of health care applications of large language models: a systematic review. Raji, I. D., Daneshjou, R. & Alsentzer, E. It's time to bench the medical exam benchmark. Han, R. et al. Randomised controlled trials evaluating artificial intelligence in clinical practice: a scoping review. Wan, P. et al. Outpatient reception via collaboration between nurses and a large language model: a randomized controlled trial. The FAIIR conversational AI agent assistant for youth mental health service provision. Li, Y. et al. ChatDoctor: a medical chat model fine-tuned on a large language model Meta-AI (LLaMA) using medical domain knowledge. Towards accurate differential diagnosis with large language models. Li, S. et al. A community co-designed LLM-powered chatbot for primary care: a randomized controlled trial. Physician perspectives on ambient AI scribes. Irving, G. et al. International variations in primary care physician consultation time: a systematic review of 67 countries. Ong, L. M. L., de Haes, J. C. J. M., Hoos, A. M. & Lammes, F. B. Doctor-patient communication: a review of the literature. Dennstädt, F., Hastings, J., Putora, P. M., Schmerder, M. & Cihoric, N. Implementing large language models in healthcare while balancing control, collaboration, costs and security. Rajkomar, A., Hardt, M., Howell, M. D., Corrado, G. & Chin, M. H. Ensuring fairness in machine learning to advance health equity. Obermeyer, Z., Powers, B., Vogeli, C. & Mullainathan, S. Dissecting racial bias in an algorithm used to manage the health of populations. Enhancing diagnostic capability with multi-agents conversational large language models. Bedi, S., Jain, S. S. & Shah, N. H. Evaluating the clinical benefits of LLMs. & Morewedge, C. K. Resistance to medical artificial intelligence. & Zhang, Z. Analyzing the medical record homepages quality in a Chinese EMR system. & Abdulnour, R. E. Conversations on reasoning: large language models in diagnosis. Cabral, S. et al. Clinical reasoning of a generative artificial intelligence model compared with physicians. Han, S. Facilitating patient portal shared access to support age-friendly health care. Liu, P. et al. Pre-train, prompt, and predict: a systematic survey of prompting methods in natural language processing. Ji, S., Pan, S., Cambria, E., Marttinen, P. & Yu, P. S. A survey on knowledge graphs: representation, acquisition and applications. Rasmy, L., Xiang, Y., Xie, Z., Tao, C. & Zhi, D. Med-BERT: pre-trained contextualized embeddings on large-scale structured electronic health records for disease prediction. & Chen, D. SimCSE: Simple contrastive learning of sentence embeddings. 2021 Conference on Empirical Methods in Natural Language Processing 6894–6910 (Association for Computational Linguistics, 2021). is supported by the National Natural Science Foundation of China (no. is supported by the National Natural Science Foundation of China (no. 82260008) and the 2023 Bagui Young Top Talents Training Project of Guangxi Zhuang Autonomous Region. is supported by the National Natural Science Foundation of China (no.62532001). These authors contributed equally: Xinge Tao, Shuya Zhou, Kai Ding. School of Population Medicine and Public Health, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, China Xinge Tao, Shuya Zhou, Sairan Li, Boyou Wu, Qirui Huang, Wangyue Chen, Muzi Shen, En Meng & Shasha Han Key Laboratory of Basic Research in Respiratory Diseases, Health Commission of Guangxi Zhuang Autonomous Region, Guilin, China Key Laboratory of Respiratory Diseases, Education Department of Guangxi Zhuang Autonomous Region, Guilin, China Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai, China Wangxuan Institute of Computer Technology, Peking University, Beijing, China Department of Information Technology, The First Affiliated Hospital of Guilin Medical University, Guilin, China Pattern Recognition Center, WeChat AI, Tencent Inc, Beijing, China State Key Laboratory of Respiratory Health and Multimorbidity, Beijing, China Key Laboratory of Pathogen Infection Prevention and Control (Peking Union Medical College), Ministry of Education, Beijing, China Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar All authors contributed to the study design. Correspondence to Libing Ma or Shasha Han. are employees of WeChat AI, Tencent, which provided computational resources and technical support for this research. The other authors declare no competing interests. Nature Medicine thanks Guosheng Yin and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Lorenzo Righetto, in collaboration with the Nature Medicine editorial team. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. a, The co-designed architecture and clinical integration of PreA. Patients first interact with the chatbot, which generates a structured referral report for the specialist to review via the clinician interface prior to standard consultation. The first cycle involved adversarial testing with community and clinical stakeholders. The second cycle used GPT-4-powered virtual patient simulations to optimize the model against co-designed evaluation metrics. c, Experimental comparison of the co-designed PreA model against a local data-tuned counterpart, created by fine-tuning the base PreA model on local primary care dialogs. d, Multicenter randomized controlled trial design. Patients were randomized to one of three arms before specialist consultation: PreA-only (independent use of PreA), PreA-human (staff-supported use of PreA), or a No-PreA control (usual care). Box plots depict patient consultation time across the three trial arms (PreA-only, PreA-human, No-PreA). Dot plots show patient-reported experience metrics (ease of communication, perceived physician attentiveness, interpersonal regard, patient satisfaction, and future acceptability) and physician-reported perceived value on care coordination, with error bars representing standard deviation. Sample sizes for each subgroup are provided in Table 1. We assessed the normality of value distributions and used two-sample t-tests with unequal variances for intergroup comparisons. For significantly skewed dimensions, we employed non-parametric Mann-Whitney U-tests. The Benjamini-Hochberg adjustment was applied for multiple testing corrections based on the total number of tests. Box plots depict patient consultation time across the three trial arms (PreA-only, PreA-human, No-PreA). Dot plots show patient-reported experience metrics (ease of communication, perceived physician attentiveness, interpersonal regard, patient satisfaction, and future acceptability) and physician-reported perceived value on care coordination, with error bars representing standard deviation. Sample sizes for each subgroup are provided in Table 1. We assessed the normality of value distributions and used two-sample t-tests with unequal variances for intergroup comparisons. For significantly skewed dimensions, we employed non-parametric Mann-Whitney U-tests. The Benjamini-Hochberg adjustment was applied for multiple testing corrections based on the total number of tests. Box plots depict patient consultation time across the three trial arms (PreA-only, PreA-human, No-PreA). Dot plots show patient-reported experience metrics (ease of communication, perceived physician attentiveness, interpersonal regard, patient satisfaction, and future acceptability) and physician-reported perceived value on care coordination, with error bars representing standard deviation. Sample sizes for each subgroup are provided in Table 1. We assessed the normality of value distributions and used two-sample t-tests with unequal variances for intergroup comparisons. For significantly skewed dimensions, we employed non-parametric Mann-Whitney U-tests. The Benjamini-Hochberg adjustment was applied for multiple testing corrections based on the total number of tests. Box plots depict patient consultation time across the three trial arms (PreA-only, PreA-human, No-PreA). Dot plots show patient-reported experience metrics (ease of communication, perceived physician attentiveness, interpersonal regard, patient satisfaction, and future acceptability) and physician-reported perceived value on care coordination, with error bars representing standard deviation. Sample sizes for each subgroup are provided in Table 1. We assessed the normality of value distributions and used two-sample t-tests with unequal variances for intergroup comparisons. For significantly skewed dimensions, we employed non-parametric Mann-Whitney U-tests. The Benjamini-Hochberg adjustment was applied for multiple testing corrections based on the total number of tests. Study protocol and statistical analysis plan (English). Study protocol and statistical analysis plan (Chinese). Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Tao, X., Zhou, S., Ding, K. et al. An LLM chatbot to facilitate primary-to-specialist care transitions: a randomized controlled trial. Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Sign up for the Nature Briefing: Translational Research newsletter — top stories in biotechnology, drug discovery and pharma.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Nature Synthesis (2026)Cite this article C-heteroaryl glycosides, predominantly in unprotected forms, are common entities in bioactive molecules and have extensive applications in chemistry and biology. However, the chemical synthesis of these glycosides remains challenging owing to the lack of methods that directly leverage naturally occurring (native) sugars as substrates. Here we show that fully unprotected native sugars, capped as redox-active glycosyl sulfide donors, can be merged with N-heteroarenes in the presence of triethylamine and a photocatalyst under mild visible-light irradiation. The C–C coupling transformation proceeds with control over chemo-, site- and stereoselectivities and is compatible with a diverse range of N-heteroarenes bearing acidic and basic functional groups. The utility of this method is highlighted by the glycosylation of nucleosides, as well as by the direct coupling of d-mannose with pentoxifylline to generate a compound exhibiting glycogen-metabolism-inhibitory properties. In contrast to previously established mechanisms, the photocatalytic species is found to trigger the in situ generation of a thiyl radical that promotes hydrogen atom transfer to afford the target product, with triethylamine serving as a reductant through photoinduced charge-transfer complexation with the glycosyl sulfide. This is a preview of subscription content, access via your institution Subscribe to this journal Receive 12 digital issues and online access to articles $119.00 per year only $9.92 per issue Buy this article Prices may be subject to local taxes which are calculated during checkout Crystallographic data are available free of charge from the Cambridge Crystallographic Data Centre under deposition numbers CCDC-2449025 (46) and CCDC-2449026 (31). All other data are available in the Article or its Supplementary Information. Bokor, E. et al. C-glycopyranosyl arenes and hetarenes: Synthetic methods and bioactivity focused on antidiabetic potential. Article PubMed Google Scholar Yang, Y. & Yu, B. Recent advances in the chemical synthesis of C-glycosides. Article PubMed Google Scholar Báti, G., He, J.-X., Pal, K. B. & Liu, X.-W. Stereo- and regioselective glycosylation with protection-less sugar derivatives: an alluring strategy to access glycans and natural products. Article PubMed Google Scholar Štambaský, J., Hocek, M. & Kočovský, P. C-nucleosides: synthetic strategies and biological applications. Article PubMed Google Scholar Villadsen, K., Martos-Maldonado, M. C., Jensen, K. J. & Thygesen, M. B. Chemoselective reactions for the synthesis of glycoconjugates from unprotected carbohydrates. Article PubMed Google Scholar Cao, X. et al. Carbohydrate-based drugs launched during 2000-2021. Google Scholar Warren, T. K. et al. Therapeutic efficacy of the small molecule GS-5734 against ebola virus in rhesus monkeys. Google Scholar Kun, S. et al. New synthesis of 3-(β-d-glucopyranosyl)-5-substituted-1,2,4-triazoles, nanomolar inhibitors of glycogen phosphorylase. Article PubMed Google Scholar Ősz, E. et al. Preparation of 3,5-bis-(β-d-glycopyranosyl)-1,2,4-thiadiazoles from C-(β-d-glycopyranosyl)thioformamides. Google Scholar Cohn, W. E. Pseudouridine, a carbon-carbon linked ribonucleoside in ribonucleic acids: isolation, structure, and chemical characteristics. Article PubMed Google Scholar Wang, M. et al. Remdesivir and chloroquine effectively inhibit the recently emerged novel coronavirus (2019-nCoV) in vitro. Google Scholar Chen, A., Yang, B., Zhou, Z. & Zhu, F. Recent advances in transition-metal-catalyzed glycosyl cross-coupling reactions. Chem Catal. Google Scholar Xia, L., Fan, W., Yuan, X.-A. & Yu, S. Photoredox-catalyzed stereoselective synthesis of C-nucleoside analogues from glycosyl bromides and heteroarenes. Google Scholar Wang, Q. et al. Palladium-catalysed C−H glycosylation for synthesis of C-aryl glycosides. Google Scholar Pöhlmann, J., Yuan, B., Purushothaman, R., Wu, J. & Ackermann, L. Photoinduced ruthenium-catalyzed meta-C–H glycosylation. Google Scholar Xu, S. et al. Generation and use of glycosyl radicals under acidic conditions: glycosyl sulfinates as precursors. Google Scholar Liu, Y., Wu, J., Zhou, L., Wei, C. & Chen, H. Synthesis of N-heteroaryl C-glycosides and polyhydroxylated alkanes with diaryl groups from unprotected sugars. ACS Omega. Google Scholar Obradors, C. et al. Direct and catalytic C-glycosylation of arenes: Expeditious synthesis of the remdesivir nucleoside. Google Scholar Jiang, Y. et al. Direct radical functionalization of native sugars. Google Scholar Cao, S. et al. Regiodivergent functionalization of protected and unprotected carbohydrates using photoactive 4-tetrafluoropyridinylthio fragment as an adaptive activating group. Google Scholar Li, S. et al. Iridium photoredox-catalyzed stereoselective C-glycosylation with tetrafluoropyridin-4-yl thioglycosides: a facile synthesis of C-α/β-glucogallins and their antioxidant activity. Google Scholar Dewanji, A. et al. A general arene C–H functionalization strategy via electron donor–acceptor complex photoactivation. Article PubMed Google Scholar Xie, D., Zeng, W., Yang, J. & Ma, X. Visible-light-promoted direct desulfurization of glycosyl thiols to access C-glycosides. Google Scholar Wang, Q. et al. Visible light activation enables desulfonylative cross-coupling of glycosyl sulfones. Google Scholar Shang, W. & Niu, D. Radical pathway glycosylation empowered by bench-stable glycosyl donors. Article PubMed Google Scholar Xu, L.-Y., Fan, N.-L. & Hu, X.-G. Recent development in the synthesis of C-glycosides involving glycosyl radicals. Article PubMed Google Scholar Chen, A., Cheng, G. & Zhu, F. Recent advances in stereoselective synthesis of non-classical glycosides. Tetrahedron Chem. et al. Visible-light-driven synthesis of N-heteroaromatic carboxylic acids by thiolate-catalysed carboxylation of C(sp2)–H bonds using CO2. Capaldo, L., Ravelli, D. & Fagnoni, M. Direct photocatalyzed hydrogen atom transfer (HAT) for aliphatic C–H bonds elaboration. Yoshida, N. et al. Direct dehydrative pyridylthio-glycosidation of unprotected sugars in aqueous media using 2-chloro-1,3-dimethylimidazolinium chloride as a condensing agent. Asian J. Bryden, M. A. & Zysman-Colman, E. Organic thermally activated delayed fluorescence (TADF) compounds used in photocatalysis. Spell, M. L. et al. A visible-light-promoted O-glycosylation with a thioglycoside donor. Abe, H., Shuto, S. & Matsuda, A. Highly α- and β-selective radical C-glycosylation reactions using a controlling anomeric effect based on the conformational restriction strategy. A study on the conformation−anomeric effect−stereoselectivity relationship in anomeric radical reactions. & Dupuis, J. Diastereoselective syntheses of C-glycopyranosides. Proctor, R. S. J., Davis, H. J. & Phipps, R. J. Catalytic enantioselective minisci-type addition to heteroarenes. Fuse, H. et al. Photocatalytic redox-neutral hydroxyalkylation of N-heteroaromatics with aldehydes. Jin, J. & MacMillan, D. W. C. Alcohols as alkylating agents in heteroarene C–H functionalization. Shi, Q. et al. Single-electron oxidation-initiated enantioselective hydrosulfonylation of olefins enabled by photoenzymatic catalysis. Xuan, L. et al. Remote C(sp3)–H heteroarylation of N-fluorocarboxamides with quinoxalin-2(1H)-ones under visible-light-induced photocatalyst-free conditions. Green Chem. Dong, J. et al. Formyl-selective deuteration of aldehydes with D2O via synergistic organic and photoredox catalysis. Lahdenperä, A. S. K. et al. A chiral hydrogen atom abstraction catalyst for the enantioselective epimerization of meso-diols. Guo, S. Y. et al. Visible light-induced chemoselective 1,2-diheteroarylation of alkenes. Zhang, C. et al. Visible-light-induced synthesis of heteroaryl C-glycosides via decarboxylative C–H glycosylation. Adak, L. et al. Synthesis of aryl C-glycosides via iron-catalyzed cross coupling of halosugars: stereoselective anomeric arylation of glycosyl radicals. Chen, H., Fan, W., Yuan, X. & Yu, S. Site-selective remote C(sp3)–H heteroarylation of amides via organic photoredox catalysis. Le Saux, E., Georgiou, E., Dmitriev, I. A., Hartley, W. C. & Melchiorre, P. Photochemical organocatalytic functionalization of pyridines via pyridinyl radicals. Van Aerschot, A. A. et al. Antiviral activity of C-alkylated purine nucleosides obtained by cross-coupling with tetraalkyltin reagents. Chang, L. C. W. et al. 2,6-disubstituted and 2,6,8-trisubstituted purines as adenosine receptor antagonists. Gillingham, D., Geigle, S. & Anatole von Lilienfeld, O. Properties and reactivity of nucleic acids relevant to epigenomics, transcriptomics, and therapeutics. Xie, R. et al. Late-stage guanine C8-H alkylation of nucleosides, nucleotides, and oligonucleotides via photo-mediated minisci reaction. Chrysina, E. D. et al. Kinetic and crystallographic studies on 2-(β-d-glucopyranosyl)-5-methyl-1, 3, 4-oxadiazole, -benzothiazole, and -benzimidazole, inhibitors of muscle glycogen phosphorylase b. Evidence for a new binding site. Protein Sci. Download references This research is supported by the Ministry of Education of Singapore Academic Research Fund Tier 2: A-8002999-00-00 (M.J.K. ), the National Research Foundation, Prime Minister's Office, Singapore under its Campus for Research Excellence and Technological Enterprise (CREATE) SM3 programme and A*STAR under its Manufacturing, Trade and Connectivity (MTC) Programmatic Fund: M25O1b0015 (M.J.K.). We thank I. I. Roslan (National University of Singapore) for the X-ray crystallographic measurements. These authors jointly supervised this work: Eric Chun Yong Chan, Ming Joo Koh. Department of Chemistry, National University of Singapore, Singapore, Republic of Singapore Qian-Yi Zhou, Jun Wu, Songge Li & Ming Joo Koh Department of Pharmacy and Pharmaceutical Sciences, National University of Singapore, Singapore, Republic of Singapore Daniel Zhi Wei Ng, Wei Liang Leow & Eric Chun Yong Chan Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar conceived the work. conducted the optimization and reaction scope studies. conducted the mechanistic studies. conducted the biological studies under the direction of E.C.Y.C. directed the research. All authors contributed to the writing of the manuscript. Correspondence to Eric Chun Yong Chan or Ming Joo Koh. The authors declare no competing interests. Nature Synthesis thanks Yong-Min Liang, Feng Zhu and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Joel Cejas-Sánchez, in collaboration with the Nature Synthesis team. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Sections 1–11, including Supplementary Figs. 1–13, Discussion and Tables 1–9. Crystallographic data for compound 31, CCDC 2449026. Crystallographic data for compound 46, CCDC 2449025. The raw data of cyclic voltammograms of Supplementary Fig. The raw data of bioactivity investigation of Supplementary Fig. The raw data of UV–vis spectroscopy studies of Supplementary Fig. The raw data of the quenching ability of NEt3 of Supplementary Fig. The raw data of the quenching ability of isoquinoline of Supplementary Fig. The raw data of the quenching ability of the glycosyl donor of Supplementary Fig. The raw data of the quenching ability of NEt3 and thiol of Supplementary Fig. Statistical source data for Fig. Statistical source data for Fig. Statistical source data for Fig. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. Reprints and permissions Zhou, QY., Ng, D.Z.W., Wu, J. et al. Photocatalytic coupling of unprotected sugars and N-heteroarenes. Received: 19 August 2025 Published: 19 January 2026 Version of record: 19 January 2026 DOI: https://doi.org/10.1038/s44160-025-00980-8 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative ISSN 2731-0582 (online) © 2026 Springer Nature Limited Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.