MedGemma:开放的医疗基础模型支持临床AI应用
Open foundation models for clinical AI
Andrew Sellergren及其团队开发了MedGemma,这是一种开放的医疗基础模型家族,包括多模态模型和医学图像编码器,适用于广泛的医疗AI应用。在多个基准测试中,MedGemma表现出色,特别是在处理跨患者记录的信息合成方面。此外,MedGemma 4B模型仅用10%的任务特定训练数据就获得了优于Gemma 3模型的性能,显示出其在数据高效适应方面的潜力。
- Research Highlight
- Published: 12 November 2026
Medical AI
Nature Computational Science (2026) Cite this article
Large language models have increasingly demonstrated strong medical abilities, while vision–language models extend these capabilities to clinical images and other multimodal data. Yet, strong performance on individual benchmarks does not automatically translate into useful clinical applications. To be deployed in healthcare, models must adapt to local settings, integrate with clinical information systems, and support complex reasoning across heterogeneous patient data. In order to address this challenge, Andrew Sellergren, Lin Yang, and colleagues introduce MedGemma, an open family of medical foundation models designed to support a broad spectrum of healthcare artificial intelligence (AI) applications.
MedGemma includes multimodal models with 4B and 27B parameters, a text-only 27B parameter model, and MedSigLIP, a medical image encoder. The models were developed using medical text and images spanning radiology, pathology, dermatology and ophthalmology. Across a range of benchmarks, MedGemma consistently outperformed its Gemma 3 base models. In AgentClinic, a simulated environment designed to evaluate longitudinal clinical reasoning, MedGemma 27B showed strong performance and retained its capabilities when evaluated on out-of-distribution MIMIC-IV cases. Post-training on synthetic longitudinal electronic health records (EHRs) improved performance particularly on questions requiring synthesis of information across interdependent patient records. Moreover, fine-tuning MedGemma 4B with only 10% of the task-specific training data yielded stronger performance than similarly fine-tuning the corresponding Gemma 3 model, highlighting its potential for data-efficient adaptation.
This is a preview of subscription content, access via your institution
Access options
Access through your institution
Access Nature and 54 other Nature Portfolio journals
Get Nature+, our best-value online-access subscription
32.70 € / 30 days cancel any time
Subscribe to this journal
Receive 12 digital issues and online access to articles
111.21 € per year only 9.27 € per issue
Buy this article
- Purchase on SpringerLink
- Instant access to the full article PDF.
39,95 €
Prices may be subject to local taxes which are calculated during checkout
Additional access options:
Author information
Authors and Affiliations
Nature Computational Science https://www.nature.com/natcomputsci/
Vikas Bansal
Authors
- Vikas Bansal
Corresponding author
Correspondence to Vikas Bansal.
About this article
Cite this article
Bansal, V. Open foundation models for clinical AI. Nat Comput Sci (2026). https://doi.org/10.1038/s43588-026-01094-7
Published: 12 November 2026
Version of record: 12 November 2026
DOI: https://doi.org/10.1038/s43588-026-01094-7
来源:Nature 健康护理主题 · nature.com