人工智能和机器学习在精准医疗中的新兴应用
Emerging applications of machine learning and AI for predictive modeling in precision medicine
文章总结了31篇论文,强调AI在精准医疗中的应用需结合生物知识,实现可解释性、人性化设计和公平性。视频分析、信号分析、可穿戴设备和多组学集成展示了AI在诊断、预测和干预方面的进步。
Pure data-driven artificial intelligence (AI) is insufficient for clinical adoption. This editorial summarizes 31 papers showing that precision medicine requires integrating structural domain knowledge before model training. We identify three necessary transitions: anchoring algorithms in biological laws for mechanistic interpretability; designing collaborative, human-centered clinical interfaces; and scaling equitable, population-level interventions. These shifts advance medical AI from technical novelty toward trustworthy, clinically grounded, and ethically deployed foundational science.
Introduction
This special collection highlights 31 works across video analysis, signal analysis, digital health integration, and population health, arguing that medical AI’s future lies in minimalistic, interpretable tools combining domain knowledge with state-of-the-art algorithms to improve patient outcomes and equity.
Across the four thematic areas represented in this collection, a consistent narrative emerges. Video-based diagnostics are transitioning from controlled laboratory settings to portable, accessible tools capable of specialist-level accuracy in everyday environments. Signal analysis is pivoting toward lightweight devices performing complex tasks in natural environments where traditional methods struggle to balance precision with ecological validity. Wearable and biomarker integration now capture dynamic health complexity through multivariate models that better reflect real-world physiology than isolated metrics. Finally, population health applications are demonstrating that AI’s ultimate test lies in whether it can deliver equitable benefit at scale (Fig. 1). Together, these threads reveal a coherent trajectory: medical AI is maturing from an engineering problem into a clinical science, where trust, interpretability, and equity are not afterthoughts but foundational design principles.
Video analysis and pose estimation
Laboratory-based video analysis is precise but expensive, limiting access to specialized care. New paradigms focus on small, mobile hardware suitable for any environment, including low-resource settings. One of the most important emerging fields has been video-based tools helping physicians classify pathologies or provide objective second opinions. Smartphone-based graph convolutional networks now assess Parkinsonian gait asymmetry at home1, while VideoPose3D extracts accurate 3D joint angles by removing noise via clustering2. Open-source models transform raw video into interpretable Autism Spectrum Disorder metrics3, and smartphone systems combine video analysis with large language models for ophthalmologic evaluation4. Deep learning also enables real-time ultrasound lesion detection5, noninvasive blood count estimation from ocular videos6, and cardiovascular risk prediction from echocardiography7. Collectively, these advances enable automated, objective assessments, shifting video diagnostics toward portable, lightweight alternatives that achieve specialist-level accuracy beyond the lab.
Signal analysis for individual prediction
While video analysis captures external movement and physical presentation, signal analysis examines the body’s internal physiological waveforms, extending precision medicine from observation to continuous internal monitoring. Signal analysis is pivoting toward lightweight devices performing complex tasks in natural environments where traditional methods struggle to balance precision with ecological validity. Personalized Deep Brain Stimulation can now be optimized during sleep at home8, and multicenter training harmonizes variables to predict Parkinson’s cognitive decline without overfitting9. Stroke recurrence prediction improves by masking acute lesions to derive brain age from healthy tissue alone10, while Graph Neural Networks identify network disruptions linked to depression treatment remission11. In cardiology, deep learning detects subclinical heart failure from ECGs12, contrastive multimodal learning bypasses labeled dataset bottlenecks13, and interpretable transformers align explanations with clinical reasoning14. Transformer-based models further improve outcome predictions from routine records with reduced bias15.
A Video-based motion capture and pose estimation for biomechanical assessment; B multimodal signal analysis including electrophysiological recordings, statistical modeling, and pattern recognition; and C continuous monitoring through wearable devices integrated with genomic biomarkers and electronic health records. D Convergence of these heterogeneous data streams through machine learning frameworks to enable individualized clinical decision support, collaborative care coordination, and scalable population health interventions, advancing equitable healthcare delivery across diverse patient populations.
Wearables, biomarkers, and digital health records
Building on isolated physiological signals, the integration of continuous data streams with clinical records marks a significant maturation of the field. Multivariate models integrating wearables, biomarkers, and electronic health records (EHR) now capture dynamic health complexity better than isolated metrics, supporting preventive care through comprehensive continuous monitoring. Retrieval-Augmented Generation agents provide verified, personalized postoperative support via platforms like WeChat16. Multi-omics advances include blood transcriptome models for pre-symptomatic Alzheimer’s based on biological causality17, metabolic pathway identification in pancreatitis18, macrophage subpopulation discovery in liver disease19, frailty subtype optimization20, and metagenomic differentiation of lung cancer versus infection21. Explainable AI improves acceptance in these contexts. LLM embeddings enhance EHR-based pancreatic cancer prediction to 86%22, while antibiotic prescribing models combine probabilistic predictions with physician preferences23.
Applications and future directions
Advances into clinical practice require a deliberate shift from model development to implementation. Academia must steer AI toward population health impact through three foundational shifts. First, mechanistic models anchored in biological laws restore interpretability and safety24,25, while virtual cell models simulate preclinical drug behavior26. Pure data-driven AI is insufficient because data alone cannot explain system behavior; hybrid models yield scientifically verifiable predictions rather than black boxes. Second, self-evolving multi-agent systems structure clinical decision-making27, but explanation design must align with human diagnostic reasoning rather than abstract mathematical weights28. Third, AI enables population-level action by automating outreach based on continuous data streams29, requiring deliberate navigation of ethical risks; incorporating social determinants reduces inequality30 and enables scalable screening in underserved communities31.
Conclusion
The most impactful applications of medical AI are those grounded in biological causality and physiological laws. This collection demonstrates that domain knowledge is not a supplement to machine learning, but its critical substrate. Mechanistic interpretability, human-centered collaboration, and equitable deployment form a necessary sequence where trust enables scale and equity fulfills medicine’s ethical imperative. Challenges including prospective validation, adaptive regulatory frameworks, and equitable distribution require interdisciplinary rigor beyond algorithms. We invite readers to engage with this issue as both a record and a provocation for AI’s role as a foundational scientific tool.
Data availability
No datasets were generated or analysed during the current study.
References
Han, J. et al. Deep learning-enabled accurate assessment of gait impairments in Parkinson’s disease using smartphone videos. npj Digit. Med. 9, 98 https://doi.org/10.1038/s41746-025-02150-8 (2026).
Natraj, S. et al. 3D pose estimation for scalable remote gait kinematics assessment. npj Digit. Med. 9, 37 https://doi.org/10.1038/s41746-025-02211-y (2026).
Kim, D. Y. et al. Automated AI based identification of autism spectrum disorder from home videos. npj Digit. Med. 8, 607 https://doi.org/10.1038/s41746-025-01993-5 (2025).
Li, H. et al. From blink to care: smartphone video–based functional analysis in pediatric blepharoptosis. npj Digit. Med. 9, 358 https://doi.org/10.1038/s41746-026-02510-y (2026).
Gu, J. et al. Automated deep learning for real-time focal liver lesions detection in ultrasound videos a multicenter study. npj Digit. Med. https://doi.org/10.1038/s41746-026-02629-y (2026).
Denis, T. et al. Towards noninvasive blood count using a deep learning pipeline from bulbar conjunctiva videos. npj Digit. Med. https://doi.org/10.1038/s41746-026-02598-2 (2026).
Rawlani, M. et al. Artificial intelligence prediction of age from echocardiography as a marker for cardiovascular disease. npj Digit. Med. 8, 688 https://doi.org/10.1038/s41746-025-02050-x (2025).
Smyth, C. et al. Personalized supervised and unsupervised intracranial sleep decoding during deep brain stimulation. npj Digit. Med. https://doi.org/10.1038/s41746-026-02368-0 (2026).
Loo, R. T. J. et al. Multi-cohort machine learning identifies predictors of cognitive impairment in Parkinson’s disease. npj Digit. Med. 8, 482 https://doi.org/10.1038/s41746-025-01862-1 (2025).
Zhou, H. et al. Deep learning-based brain age predicts stroke recurrence in acute ischemic cerebrovascular disease. npj Digit. Med. 8, 750 https://doi.org/10.1038/s41746-025-02161-5 (2025).
Liu, R. et al. Predicting antidepressant response via local-global graph neural network and neuroimaging biomarkers. npj Digit. Med. 8, 515 https://doi.org/10.1038/s41746-025-01912-8 (2025).
Stenhede, E. et al. Heart failure detection in electrocardiograms using Artificial Intelligence and pragmatic labelling. npj Digit. Med. https://doi.org/10.1038/s41746-026-02774-4 (2026).
Zhou, X. et al. Diagnosis of cardiac conditions from 12-lead ECG through natural language supervision. npj Digit. Med. 8, 697 https://doi.org/10.1038/s41746-025-02074-3 (2025).
Yisimitila, T. et al. Bridging clinical knowledge and AI: an interpretable transformer framework for ECG diagnosis. npj Digit. Med. 9, 41 https://doi.org/10.1038/s41746-025-02215-8 (2026).
Rao, S. et al. A transformer-based survival model for prediction of all-cause mortality in patients with heart failure: a multi-cohort study. npj Digit Med. 9, 118 https://doi.org/10.1038/s41746-025-02296-5 (2026).
Li, J. et al. A randomized controlled trial of a WeChat-based artificial intelligence agent for postoperative care in orthopedic patients. npj Digit. Med. 9, 105 https://doi.org/10.1038/s41746-025-02269-8 (2026).
Wu, J. et al. Multi-omics dissection of SNP-mediated immunometabolic signatures in Alzheimer’s disease reveals a novel individual predictive model. npj Digit. Med. 9, 81 https://doi.org/10.1038/s41746-025-02265-y (2026).
Deng, D. et al. Non-invasive liquid biopsy based on metabolomic profiling improves diagnosis and early warning of severe acute pancreatitis. npj Digit. Med. 9, 116 https://doi.org/10.1038/s41746-025-02294-7 (2026).
Wen, W. et al. Integrating multi-omics and machine learning systematically deciphers cellular heterogeneity and fibrotic regulatory networks in the progression from MASLD to MASH. npj Digit. Med. 9, 167 https://doi.org/10.1038/s41746-026-02352-8 (2026).
Xiao, L. et al. Metabolomic characterization of frailty identifies subtype-specific management strategies. npj Digit. Med. 8, 746 https://doi.org/10.1038/s41746-025-02075-2 (2025).
Han, D. et al. Metagenomic fingerprints in bronchoalveolar lavage differentiate pulmonary diseases. npj Digit. Med. 8, 599 https://doi.org/10.1038/s41746-025-01977-5 (2025).
Park, J. et al. Enhancing EHR-based pancreatic cancer prediction with LLM-derived embeddings. npj Digit. Med. 8, 465 https://doi.org/10.1038/s41746-025-01869-8 (2025).
Howard, A. et al. Algorithmic antibiotic decision-making in urinary tract infection using prescriber-informed prediction of treatment utility. npj Digit. Med. 9, 136 https://doi.org/10.1038/s41746-026-02369-z (2026).
Emmert-Streib, F. et al. The role of digital twins in P4 medicine: a paradigm for modern healthcare. npj Digit. Med. 8, 735 https://doi.org/10.1038/s41746-025-02115-x (2025).
Coveney, P. et al. Digital twins and Big AI: the future of truly individualised healthcare. npj Digit. Med. 8, 494 https://doi.org/10.1038/s41746-025-01874-x (2025).
Ma, C. et al. AI-driven virtual cell models in preclinical research. npj Digit Med 9, 25 https://doi.org/10.1038/s41746-025-02198-6 (2026).
Liu, Q. et al. EvoMDT: a self-evolving multi-agent system for structured clinical decision-making in multi-cancer. npj Digit. Med. 9, 124 https://doi.org/10.1038/s41746-025-02304-8 (2026).
Hur, S. et al. Comparison of SHAP and clinician friendly explanations reveals effects on clinical decision behaviour. npj Digit. Med. 8, 578 https://doi.org/10.1038/s41746-025-01958-8 (2025).
Basu, S. et al. Optimizing AI solutions for population health in primary care. npj Digit. Med. 8, 434 https://doi.org/10.1038/s41746-025-01864-z (2025).
Patel, S. Y., Barnett, M. L. & Basu, S. Predicting quality measure completion among 14 million low-income Medicaid patients. npj Digit. Med. 8, 393 https://doi.org/10.1038/s41746-025-01797-7 (2025).
Ge, M., Wang, Y. & Xu, S. From retina to brain: how deep learning closes the gap in silent stroke screening. npj Digit. Med. 8, 655 https://doi.org/10.1038/s41746-025-02021-2 (2025).
Funding
M.M. received funding from the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) Project-ID 424778381-TRR 295 (C05 and INF).
Author information
Authors and Affiliations
Chair of Informatics for Medical Technology (CIMT), Institute of Computer Science, University of Augsburg, Augsburg, Germany
Jaime Zorraquin, Manuel Bange & Muthuraman Muthuraman
Department of Dermatology, Harvard Medical School; Senior Advisor of Virtual Care, Mass General Brigham, Boston, MA, USA
Joseph C. Kvedar
Authors
- Jaime Zorraquin
- Joseph C. Kvedar
- Manuel Bange
- Muthuraman Muthuraman
Contributions
J.Z. wrote the original draft of the manuscript. J.C.K., M.B., and M.M. reviewed and edited the manuscript. All authors read and approved the final version.
Corresponding authors
Correspondence to Manuel Bange or Muthuraman Muthuraman.
Ethics declarations
Competing interests
The authors declare no competing interests.
Rights and permissions
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/.
About this article
Cite this article
Zorraquin, J., Kvedar, J.C., Bange, M. et al. Emerging applications of machine learning and AI for predictive modeling in precision medicine. npj Digit. Med. 9, 842 (2026). https://doi.org/10.1038/s41746-026-03393-9
Received: 14 August 2026
Accepted: 30 September 2026
Published: 03 November 2026
Version of record: 03 November 2026
DOI: https://doi.org/10.1038/s41746-026-03393-9
来源:Nature 健康护理主题 · nature.com