Google says health institutions and developers in several countries are using its MedGemma models to build tools for medical triage, diagnostic testing, and clinical decision support, with some applications operating offline and retaining patient data locally. Google says the models are open-weight, allowing them to be adapted to local languages and healthcare needs.
MedGemma is a collection of models built on Gemma and designed to understand medical text and images, including clinical notes, X-rays, and computed tomography scans. According to Google, the models have been downloaded more than 10 million times and used in thousands of research adaptations worldwide. The company provides the models, code, and documentation through the Health AI Developer Foundations website.
Uses in Primary Care and Remote Areas
In Uganda, Crane AI is developing the EaseHealth application using a modified version of MedGemma that runs on the device, helping healthcare workers assess symptoms, review guidelines, and make initial triage decisions without an internet connection.
In Zambia, Dawa Health uses the DawaMom and MedSigLIP applications to screen for cervical cancer. Google says the application has screened more than 3,500 women, with plans to expand within Zambia and neighboring regions. In India, Visilant has integrated MedGemma into a smartphone-based imaging system that has previously screened more than 50,000 patients for cataracts and other eye conditions.
Trials Within Hospitals and National Programs
At AIIMS Delhi, IndusDerma is being tested to assist with the triage of dermatological cases, with a design that takes into account the needs of India’s population and skin tones. The Aarogyam application also focuses on triaging outpatients and reducing the waiting period before seeing a specialist. After completing trials and clinical validation, AIIMS aims to expand these applications across its hospital network, which receives up to 15,000 outpatient visits daily.
In Indonesia, the Ministry of Health is developing a tuberculosis-detection model using MedGemma and MedSigLIP, trained on local chest X-ray images, to support the goal of screening 50 million people annually.
What Is Changing in Practice?
The practical value lies in making models available that can run on devices or local servers, which may help institutions that lack reliable connectivity or need to keep patient data under their control. The open-weight nature also allows tools to be adapted to specific data, languages, and health priorities instead of relying on a general cloud service.
However, these applications do not represent independent diagnostic or treatment tools. Google explains that MedGemma and MedSigLIP require validation, adaptation, and appropriate modification for each use case, and that their outputs may be inaccurate and must not be used directly to make diagnostic or patient-management decisions or treatment recommendations. Clinical trials, independent validation, and connection to the medical context therefore remain essential conditions before operational expansion.