| name | medgemma-health-ai |
| description | Build and evaluate medical text and vision applications with Google MedGemma, including MedGemma 1.5 workflows for CT, MRI, whole-slide pathology, longitudinal chest X-rays, lab reports, and EHR text. Use when prototyping, fine-tuning, deploying, or validating MedGemma-based health AI under clinical data and safety controls. |
MedGemma Health AI
Use MedGemma as a development foundation. It requires use-case-specific validation and human oversight before any clinical deployment.
Model Selection
- Use MedGemma 1.5 4B for multimodal medical text generation, 3D CT or MRI, multi-patch whole-slide pathology, longitudinal chest X-rays, bounding-box localization, lab-document extraction, and EHR understanding.
- Consider MedGemma 1 variants when a 27B text-only model is required and the use case matches its model card.
- Use MedSigLIP instead when the task is image embedding, retrieval, or classification without text generation.
Workflow
- Define the intended user, clinical setting, decision supported, failure cost, and whether the system is research-only.
- Create a data-governance plan before model access:
- de-identify protected health information;
- document data rights and retention;
- keep inference local when policy requires it;
- review Health AI Developer Foundations terms.
- Choose local Hugging Face inference for experiments, Vertex AI endpoints for scalable online service, or Vertex AI batch prediction for large offline workloads.
- Use the official preprocessing notebooks for CT, MRI, and whole-slide images. Do not send raw high-dimensional studies through a generic 2D image path.
- Establish a baseline before fine-tuning. Split evaluation by site, device, demographic group, disease prevalence, and time when applicable.
- Evaluate task accuracy plus calibration, abstention, hallucination, subgroup performance, robustness, latency, and clinician workflow impact.
- Add retrieval, structured output validation, and deterministic post-processing only when they are separately tested.
- Require human review, audit logs, rollback criteria, and post-deployment monitoring for any real clinical workflow.
Guardrails
- Do not present model output as a diagnosis or treatment order without qualified clinician review.
- Do not rely on benchmark performance as evidence of local clinical validity.
- Do not expose PHI to an unapproved service.
- Detect and report missing modalities, truncated context, unsupported file types, and preprocessing failures.
- Preserve model identifier, prompt, image preprocessing, decoding settings, and all fine-tuning data lineage.
Output Contract
Return a model card supplement for the local use case: intended use, excluded use, dataset and subgroup coverage, metrics, failure analysis, human-oversight design, privacy controls, and deployment recommendation.
Read references/operations.md for current model details, a local inference pattern, deployment paths, and authoritative sources.