From Chest Imaging to Medical Record Q&A: MedGemma Study Shows Progress in Medical Models

A chest X-ray, a small section of a pathology slide, and years of medical records contain different forms of medical clues. If a single AI system can process these data, researchers may be able to reduce the burden of building new tools for each task. The MedGemma study published in Nature Medicine on October 6 provides evaluation evidence for this development approach, but gaps remain between test scores and reliable performance in clinical practice.

MedGemma is built on Gemma 3 and trained on medical data to process images and text. This journal publication is not the model’s debut: the research team released a technical report in July 2025, with the most recent revision in April 2026. Core improvements, including chest X-ray classification, appeared in the early report; the advance this time is the formal publication of the research following peer review.

The study reports that, on “out-of-distribution” tasks for which the test datasets were not used in training or tuning, MedGemma improved chest X-ray finding classification performance by 15.5 to 18.1% over the base model and medical image question answering by 2.6 to 10%. These tests probe the model’s capabilities when faced with different data, but they do not directly establish that it will maintain the same performance at another hospital or with another patient population.

How the model interprets images is another focus of this work. Google Research explains that its image encoder, MedSigLIP, has 400 million parameters and was tuned on chest X-rays, pathology slides, skin images, and fundus images to convert images into information the model can process. The early technical report also states that, after further fine-tuning, the model could achieve performance close to specialized methods in pneumothorax and pathology slide classification, while errors in medical record information retrieval fell by 50%. These are results for specific tasks and cannot be treated as a common accuracy rate across all medical uses.

The evaluation of chest X-ray report generation is closer to a potential workflow. The study used the MIMIC-CXR dataset. Google previously explained that, in an unblinded evaluation, one U.S. board-certified radiologist judged that 81% of reports generated by MedGemma 4B were accurate enough to lead to patient management similar to that based on the original reports. This was an expert judgment about the consequences of the reports, rather than evidence from actual patient follow-up demonstrating equivalent care outcomes. The evaluator’s knowledge of the reports’ source also limits interpretation of the evidence.

Reasoning capabilities likewise need to be understood within their testing context. The 10.8% improvement in agentic tasks listed in the technical report came from simulated clinical interactions, rather than a real-world outpatient trial. Even if the model has improved in question answering and classification, that does not mean it can already safely take on patient consultations, diagnosis, or treatment decisions.

The use these results support more directly is as a starting point for medical AI research and application development. Google explicitly states that the model needs to be adapted and validated for specific uses, and that its outputs should not be used directly for clinical diagnosis or treatment decisions. For institutions preparing to adopt it, the next step must be to determine how the model makes errors with their own data, patient populations, and workflows, and whether those errors affect care. Journal publication makes its capabilities easier to scrutinize; clinical trustworthiness must be established through subsequent validation.