Summary
Google says developers and health organisations are adapting its open-weight MedGemma models for screening, triage and public-health tools. Projects described include pilots at AIIMS Delhi and screening initiatives in India, Zambia and Indonesia.
Google says health organisations and developers are adapting its open-weight MedGemma models for medical screening, clinical triage and public-health programmes. The examples range from an offline clinical-support app being built in Uganda to pilots at AIIMS Delhi and a tuberculosis detection model under development in Indonesia.
MedGemma is a family of models built on Google’s Gemma models and designed to process medical text and images. Google says the models, code and documentation are available to developers and researchers, who can adapt them for local languages and health needs.
Projects span screening, triage and public health
In rural Uganda, Crane AI is using MedGemma to build EaseHealth, an app in which an adapted model handles clinical reasoning on a mobile device without an internet connection. The app is intended to help community health workers assess symptoms, consult guidance and make triage decisions.
In Zambia, Dawa Health’s DawaMom app combines MedGemma with MedSigLIP, a model for medical text and images. Google says the offline-capable app has been used to screen more than 3,500 women for cervical cancer, with plans to expand its reach.
In India, Visilant has screened more than 50,000 patients for cataracts and other eye conditions using a smartphone-based imaging system. The organisation is incorporating MedGemma into its screening workflows and hopes it will help identify treatable eye diseases earlier. At AIIMS Delhi, clinicians are piloting IndusDerma, a dermatology screening tool designed for Indian healthcare needs and skin tones. A separate AIIMS app, Aarogyam, is focused on outpatient triage. Its clinicians aim to reduce pre-specialist waiting time by 40%; Google says broader deployment depends on successful pilots and clinical validation.
Indonesia’s Ministry of Health is developing a tuberculosis detection model using MedGemma and MedSigLIP, trained on local chest X-ray data. The intended use is to support the ministry’s goal of screening 50 million citizens annually.
A flexible foundation, not a standalone clinical decision-maker
Open weights allow developers to adapt and deploy models in different settings. Google says MedGemma can run on-site or on a cloud server, and that developers can use the model family to build mobile tools that work offline. Local deployment can give organisations more control over where patient data is held. The examples reflect different stages of development: some tools are being built or piloted, while others are already being used for screening.
Google’s guidance says MedGemma and MedSigLIP are starting points for downstream healthcare applications. Developers need to validate, adapt or meaningfully modify them for their intended use. The models’ outputs are not intended to directly guide diagnosis, patient management or treatment, and Google warns that inaccurate outputs are possible. That distinction matters when interpreting the projects: the models provide a foundation for health tools, while clinical use depends on the application’s development and validation.