README.md

metadata

title: Medical Notes Using Google MedGemma and MedASR
emoji: 🩺
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
pinned: false
short_description: Turns a raw clinical recording into a structured SOAP note.

Medical Notes Using Google MedGemma and MedASR

Pipeline: audio file -> MedASR (transcription) -> MedGemma 4B, google/medgemma-4b-it (SOAP note) -> Gradio UI.

This takes a raw clinical recording and turns it into a structured SOAP note, attributed by speaker, without inventing anything that wasn't actually said.

Not a diagnostic tool. For demonstration/research use only — don't upload real patient data.

Both models are gated on the Hugging Face Hub, so whichever environment runs this needs an HF token with access to google/medasr and google/medgemma-4b-it.

Setup (local)

python3.12 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
hf auth login   # token needs access to the gated google/medasr and google/medgemma-4b-it repos

Model weights are pulled straight from the Hub on first run (no local models/ download step needed anymore). To point at a local copy instead, set MEDASR_MODEL_ID / MEDGEMMA_MODEL_ID to a local directory.

Run (local)

python app.py

Opens a local Gradio app: pick a wav/mp3 file, click Run, see the MedASR transcript, the generated SOAP note, and per-step timing.

Deploying to Hugging Face Spaces (ZeroGPU)

1. Create a Space with SDK "Gradio" and hardware "ZeroGPU".
2. Push this repo's contents to the Space: app.py, llm.py, requirements.txt, README.md, test_audio/clips/*.wav.
3. In the Space's Settings -> Variables and secrets, add a secret HF_TOKEN set to a token with access to the gated google/medasr and google/medgemma-4b-it repos. huggingface_hub reads this automatically.
4. llm.py's run_pipeline is decorated with @spaces.GPU, so ZeroGPU attaches a GPU only for the duration of that call; both the ASR and SOAP-note models load lazily on first invocation.