MedMink — Clinical Decision Support with 7 Google Health AI Models

MedGemma Impact Challenge submission • Clinician, Patient, and EMS workflows in one stack

MedMink orchestrates seven Google Health AI (HAI-DEF) models — MedGemma, CXR Foundation, Derm Foundation, Path Foundation, TxGemma, HeAR, and MedASR — into a single emergency-medicine workflow. A ReAct-style agent decides which model to invoke, streams reasoning via SSE, and renders transparent attribution so clinicians stay in control.

Three Pillars (requested focus)

• Clinician: Case analysis + agentic clinical reasoning with 7-step streaming pipeline, quality scorecard, clinician override layer, and model attribution strip.
• Patient: 23-language patient portal with mobile-first intake, check-in, health records, messages, and post-visit companion; RTL support for Arabic/Farsi/Urdu/Hebrew.
• EMS: AI-guided EMS run-report dictation with deterministic validation and ICD-10/medical-necessity summaries (src/api/routes/ems.py, src/medgemma/ems_interviewer.py, dashboard/src/types/ems.ts).

Highlights

• 7-model orchestration: MedGemma + CXR/Derm/Path + TxGemma + HeAR + MedASR, all attributed in UI.
• Streaming everything: SSE for case analysis, agent reasoning, consensus, interview, reassessment.
• Safety-first: Three-layer medication safety (deterministic + TxGemma + MedGemma) and 14-point client-side quality scorecard.
• Multilingual access: 23 languages with locale-aware dates and RTL rendering for patient flows.
• Clinician control: Accept/reject/modify every AI recommendation; overrides never mutate AI output.

Repo Map

• Backend FastAPI: core routes in src/api (case analysis, agent, interview, EMS, charting).
• MedGemma + tools: src/medgemma (case analyzer, consensus, medication safety, EMS, TxGemma client).
• Frontend (Next.js 14, shadcn/tailwind): dashboard/src/app and components under dashboard/src/components.
• Modal deployments: modal_app.py, modal_app_multimodal.py, modal_cxr.py, modal_derm.py, modal_path.py, modal_txgemma.py, modal_hear.py, modal_asr.py.
• Docs: feature deep dive in docs/FEATURE_REVIEW.md, competition write-up in docs/COMPETITION_SUBMISSION.md, demo script in docs/DEMO_SCRIPT.md.

Quick Start

Prereqs: Python 3.11+, Node 20+, pnpm, Modal CLI (for remote models) or local GPU/CPU for MedGemma small. Docker optional.

git clone https://github.com/<your-org>/research-synthesizer.git
cd research-synthesizer
cp .env.example .env   # fill in keys and Modal endpoints
python -m venv .venv && source .venv/bin/activate
pip install -e .
pnpm install --filter dashboard

Run backend (FastAPI):

uvicorn src.api.main:app --reload --host 0.0.0.0 --port 8001

Run frontend (Next.js):

cd dashboard
pnpm dev --port 3000

Modal models (optional, recommended for 27B + vision/audio):

• Set MEDGEMMA_MODAL_URL, TXGEMMA_MODAL_URL, CXR_FOUNDATION_MODAL_URL, DERM_FOUNDATION_MODAL_URL, PATH_FOUNDATION_MODAL_URL, HEAR_MODAL_URL, MEDASR_MODAL_URL, WHISPER_MODAL_URL in .env.
• Scripts modal_*.py deploy each endpoint.

Seed/demo data (optional):

python demo/build_seed_bundle.py
python demo/seed_all.py

Three Pillars — Where to look

• Clinician: Case pipeline and agent trace in src/api/routes/case_analysis.py and src/agents/clinical_reasoning_agent.py; UI in dashboard/src/app/(clinician)/case/page.tsx with attribution strip, foundation findings, quality scorecard, and override editors.
• Patient: Mobile-first care hub and multilingual intake in dashboard/src/app/patient and i18n utilities under dashboard/src/i18n; interview dual-agent flow in src/api/routes/interview.py.
• EMS: Run-report dictation and validation in src/api/routes/ems.py; MedGemma-driven extraction and deterministic checks in src/medgemma/ems_interviewer.py; frontend types at dashboard/src/types/ems.ts.

Safety, Privacy, and What Stays Local

• Do not commit secrets: .env is ignored. Keep .env.example only.
• Data and models are ignored by default: data/, models/, demo/output/, .venv/, .next/ are in .gitignore.
• Patient/EMS text/audio stays within your deployment; Modal endpoints are yours to provision.

More Details

• Full feature breakdown: docs/FEATURE_REVIEW.md
• Competition narrative: docs/COMPETITION_SUBMISSION.md
• Demo walkthrough: docs/DEMO_SCRIPT.md

Status

Active development for the MedGemma Impact Challenge (deadline Feb 24, 2026). Public repo: https://github.com/AEYohn/MedMink. Primary maintainer: @AEYohn.

License

MIT