Oncology ICU Rounds Prep

OncoICU Rounds Studio is a clinical-prep application using Google’s MedGemma and Gemma models. It organizes synthetic patient information for oncology ICU rounds and identifies material to review before entering the unit.

It walks through a four-step flow — Case Room → Pre-round → Rounds → Debrief — built around a library of synthetic ICU cases (no PHI). You pick a teaching lens and a learner role, work up the chart yourself, then check your reasoning against the model. The project tests where MedGemma helps, where it needs guardrails, and how its output should be presented in a clinical-prep interface.

Live demo →

Why I built it

I wanted to test the MedGemma/Gemma family outside a notebook, including the latency, prompts, and interface constraints involved in a clinical use case. Rounds prep provided a specific task with enough patient context to expose those constraints.

Stack

Google MedGemma and Gemma for the model layer.
React + Vite front end.
Deployed on Cloudflare Pages.

Interface

The Case Room: a teaching-lens grid and a library of synthetic oncology ICU cases. — Case Room — pick a teaching lens and a synthetic case (or generate one) before pre-round.

The Pre-round view: patient chart with key labs, active problems, and a differentials panel. — Pre-round — build your own problem list and presentation against the chart before checking the prep deck.

What I took away

The main problems were deciding how much context to send, how much output to show, and where the model needed explicit limits. This remains an experiment with synthetic cases, not a clinical tool.

References

MedGemma — Google DeepMind’s medically-tuned Gemma variant.
MedGemma 1.5 & MedASR — next-generation medical image interpretation and speech-to-text.
Gemma 4 — the open base model family.