AI-powered neonatal risk assessment using MedGemma — built for the MedGemma Impact Challenge
A real-time clinical decision support dashboard for Neonatal Intensive Care Units (NICU) at CHU Oujda, Morocco. The system ingests free-text French medical reports, extracts 70+ structured clinical fields using MedGemma, performs risk assessment through RAG against a corpus of 1,230 real patient records, and provides explainable clinical interpretation — all running 100% locally with no cloud dependencies.
- 🧠 MedGemma-powered extraction — Parses unstructured French medical text into 70+ structured clinical fields
- 📊 Hybrid RAG retrieval — Field-based matching (60%) + embedding similarity (40%) across 1,230 real NICU cases
⚠️ Risk assessment — Mortality and complication risk scoring grounded in similar historical cases- 🔍 Full case comparison — Shows complete clinical reports of similar cases organized by medical sections
- 🏥 Clinical interpretation — AI-generated analysis explicitly referencing real similar cases
- 🌍 Bilingual — Full French / English interface (i18n)
- 🔒 Privacy-first — Everything runs locally via llama.cpp, no patient data leaves the hospital
┌─────────────────────────────────────────────────────────────┐
│ Frontend (React 19) │
│ HomePage → PatientPage → RiskCards / RAG Cases / Timeline │
└──────────────────────────┬──────────────────────────────────┘
│ REST API
┌──────────────────────────┴──────────────────────────────────┐
│ Backend (FastAPI) │
│ │
│ ┌─────────────┐ ┌──────────────┐ ┌───────────────────┐ │
│ │ LLM Service │ │ RAG Service │ │ Storage (JSON) │ │
│ │ (MedGemma) │ │ Field+Embed │ │ patients.json │ │
│ └──────┬──────┘ └──────┬───────┘ └───────────────────┘ │
│ │ │ │
└─────────┼────────────────┼────────────────────────────────────┘
│ │
┌───────▼──────┐ ┌──────▼────────┐
│ MedGemma │ │ EmbeddingGemma │
│ (llama.cpp) │ │ (llama.cpp) │
│ :8080 │ │ :8081 │
└──────────────┘ └───────────────┘
| Component | Model | Purpose |
|---|---|---|
| Extraction + Interpretation | MedGemma 4B IT (French fine-tune) | Parses reports → structured JSON, generates clinical narrative |
| Risk Scoring | RAG (field + embedding) | Retrieves similar cases, computes weighted similarity — no LLM black-box risk |
This separation ensures risk scores are auditable and explainable: clinicians can inspect exactly which historical cases drove the assessment and why.
- Node.js ≥ 18
- Python ≥ 3.10
- llama.cpp (build instructions)
- MedGemma GGUF model (French fine-tune)
- EmbeddingGemma GGUF model
git clone https://github.com/<your-username>/nicu-neonate-dashboard.git
cd nicu-neonate-dashboard/neonate-dashboard
# Frontend
npm install
# Backend
cd backend
pip install -r requirements.txt# MedGemma (French fine-tune, ~3GB for Q5_K_M)
python download.py
# EmbeddingGemma — download a GGUF embedding model
# e.g., from https://huggingface.co# Terminal 1: MedGemma (structured extraction + interpretation)
llama-server -m MedGemmaModel/medgemma-4b-it-french-medical-assistant-q5_k_m.gguf --port 8080
# Terminal 2: EmbeddingGemma (semantic similarity for RAG)
llama-server -m <embedding-gemma.gguf> --port 8081 --embeddingcd backend
python precompute_embeddings.py
# → Generates embeddings.npy (1,224 vectors) for the RAG corpus# Terminal 3: Backend
cd backend
python main.py
# → FastAPI server at http://localhost:8000
# Terminal 4: Frontend
cd neonate-dashboard
npm run dev
# → Vite dev server at http://localhost:5173Open http://localhost:5173 — submit a French medical report and watch MedGemma analyze it.
neonate-dashboard/
├── src/ # React 19 + TypeScript frontend
│ ├── components/
│ │ ├── RagSimilarCases.tsx # Full clinical reports of similar RAG cases
│ │ ├── RiskSummaryCards.tsx # Risk level cards with trends
│ │ ├── ContributingFactors.tsx
│ │ ├── TrendTimeline.tsx # Vitals timeline visualization
│ │ ├── ClinicalInterpretation.tsx
│ │ ├── SimilarCases.tsx # LLM-generated case summaries
│ │ └── NewReportModal.tsx # Report submission form
│ ├── pages/
│ │ ├── HomePage.tsx # Unit-wide patient overview
│ │ ├── PatientPage.tsx # Individual patient risk profile
│ │ └── PatientRecordPage.tsx # Structured record viewer (70+ fields)
│ ├── data/
│ │ ├── api.ts # Backend API client
│ │ └── types.ts # TypeScript interfaces
│ ├── i18n.tsx # Bilingual translations (EN/FR)
│ └── App.tsx # Router setup
│
├── backend/ # FastAPI + Python backend
│ ├── main.py # FastAPI app + CORS
│ ├── routes.py # API endpoints (/api/reports, /api/patients, /api/stats)
│ ├── llm_service.py # MedGemma integration (Gemma chat template)
│ ├── rag_service.py # Hybrid RAG: field matching + embedding similarity
│ ├── precompute_embeddings.py # One-time embedding generation script
│ ├── storage.py # JSON-based patient record persistence
│ ├── models.py # Pydantic schemas (70+ clinical fields)
│ ├── .env # Server configuration
│ ├── requirements.txt # Python dependencies (zero OpenAI)
│ └── rag_data/ # RAG corpus
│ ├── cr_1.json ... cr_1230.json # 1,230 real patient records
│ ├── embeddings.npy # Precomputed embedding vectors
│ └── embeddings_index.json # Filename-to-index mapping
1. Clinician pastes French medical report
│
2. Quick regex extraction (weight, GA, sex, CRP, APGAR)
│
3. RAG retrieval: field matching (60%) + embedding similarity (40%)
→ Top 10 similar cases from 1,230-record corpus
│
4. MedGemma prompt: system instructions + report + similar cases
→ Gemma chat template: <start_of_turn>user ... <end_of_turn>
│
5. MedGemma outputs JSON with:
• "record": 70+ structured clinical fields
• "dashboard": risk assessment, contributing factors,
clinical interpretation grounded in real similar cases
│
6. Frontend renders: risk cards, trend timeline,
full RAG case reports, clinical interpretation
The hybrid RAG engine combines two complementary approaches:
| Method | Weight | Strengths |
|---|---|---|
| Field-based | 60% | Precise clinical matching (weight ±200g, GA ±2 weeks, same sex, CRP range) |
| Embedding-based | 40% | Semantic similarity (catches diagnosis paraphrasing, related conditions) |
Field matching provides explainability (highlighted per-field comparisons in the UI), while embeddings improve recall for semantically similar but differently worded cases.
| Layer | Technology |
|---|---|
| Frontend | React 19, TypeScript, Vite, React Router |
| Backend | FastAPI, Pydantic, Uvicorn |
| LLM | MedGemma 4B IT (French fine-tune, GGUF) |
| Embeddings | EmbeddingGemma (GGUF) |
| Inference | llama.cpp (local, no cloud) |
| RAG | Field-based + cosine similarity (NumPy) |
| Data | JSON files (1,230 real NICU records) |
| i18n | Custom React Context (EN/FR) |
Zero external API dependencies. Everything runs on-premise for maximum patient data privacy.
- 100% local inference — Patient data never leaves the hospital network
- No cloud APIs — MedGemma and EmbeddingGemma run via llama.cpp on local hardware
- Explainable risk scores — RAG-based, not black-box LLM predictions
- Clinical decision support — Assists clinicians, never replaces their judgment
- Auditable — Every risk score can be traced back to specific similar cases and matching fields
The system includes a corpus of 1,230 real neonatal patient records from CHU Oujda NICU:
| Statistic | Value |
|---|---|
| Total records | 1,230 |
| Structured fields per record | 70+ |
| Embedded records | 1,224 |
| Embedding dimensions | 1,536 |
| Survival rate | ~85% |
Contributions are welcome! Areas of interest:
- Improving MedGemma prompt engineering for French medical text
- Adding new clinical sections to the RAG corpus
- Implementing embedding-based reranking
- Database migration (JSON → PostgreSQL) for production scale
- Additional language support
This project was built for the MedGemma Impact Challenge by Google Research.
Built with ❤️ for neonatal care at CHU Oujda, Morocco