🦠 EpiCast — Agentic Disease Surveillance with MedGemma

MedGemma Impact Challenge Submission

Turning frontline clinical encounters into real-time epidemiological intelligence using an agentic pipeline of HAI-DEF models — running entirely offline at the edge.

---

The Problem

Disease outbreaks in low-resource settings are detected weeks too late. Community health workers (CHWs) see early signals — unusual symptom clusters, sudden spikes in presentations — but lack tools to aggregate, analyze, and escalate. Traditional surveillance relies on manual reporting chains: paper forms → district offices → national databases. By the time data reaches epidemiologists, outbreaks have already spread.

WHO estimates that 60% of outbreaks in LMICs are detected by informal channels, not formal surveillance systems. EpiCast bridges this gap.

The Solution

EpiCast is an agentic disease surveillance system that transforms unstructured clinical encounters into structured epidemiological intelligence, running entirely on local hardware without internet connectivity.

Architecture Overview

┌─────────────────────────────────────────────────────────────────┐
│                        EpiCast Pipeline                         │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌──────────────┐    ┌───────────────────┐    ┌──────────────┐ │
│  │  MedASR       │───▶│  Intake Agent      │───▶│  Syndromic   │ │
│  │  (Speech→Text)│    │  (MedGemma 4B)     │    │  Database    │ │
│  └──────────────┘    │  Extracts:         │    └──────┬───────┘ │
│                       │  • Symptoms/Signs  │           │         │
│  ┌──────────────┐    │  • Demographics    │           │         │
│  │  MedSigLIP   │───▶│  • Severity        │           │         │
│  │  (Image Enc.) │    │  • Geo/Temporal    │           │         │
│  └──────────────┘    └───────────────────┘           │         │
│                                                        │         │
│  ┌────────────────────────────────────────────────────▼───────┐ │
│  │                Surveillance Analytics Agent                 │ │
│  │                    (MedGemma 27B)                           │ │
│  │  • Temporal aggregation of syndromic signals                │ │
│  │  • Anomaly detection (Poisson / CUSUM thresholds)           │ │
│  │  • Time-series forecasting (diffusion-based)                │ │
│  │  • Situation report generation                              │ │
│  │  • Outbreak probability scoring                             │ │
│  └────────────────────────────────────────────────────┬───────┘ │
│                                                        │         │
│  ┌────────────────────────────────────────────────────▼───────┐ │
│  │                 Escalation & Alert Agent                    │ │
│  │  • Auto-generates alerts for district health officers       │ │
│  │  • Patient-facing health advisories (multilingual)          │ │
│  │  • FHIR-compatible surveillance reports                     │ │
│  │  • Actionable recommendations with evidence                 │ │
│  └────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘

HAI-DEF Models Used

Novel Contributions

1. Syndromic Surveillance as a Novel Task: Using MedGemma not for individual diagnosis, but for population-level epidemiological signal extraction — a task it was never explicitly trained for
2. Agentic Multi-Model Orchestration: Four HAI-DEF models working as specialized agents in a coherent pipeline
3. Diffusion-Based Outbreak Forecasting: Integrating probabilistic time-series forecasting with MedGemma's clinical reasoning
4. Edge-First Design: Full intake pipeline runs on MedGemma 4B offline; syncs to 27B when connectivity is available

---

Project Structure

epicast/
├── README.md
├── requirements.txt
├── configs/
│   ├── model_config.yaml          # Model paths, LoRA config, inference params
│   └── surveillance_config.yaml    # Disease definitions, alert thresholds
├── src/
│   ├── agents/
│   │   ├── __init__.py
│   │   ├── intake_agent.py         # MedASR + MedGemma 4B syndromic extraction
│   │   ├── image_triage_agent.py   # MedSigLIP + MedGemma 4B image analysis
│   │   ├── surveillance_agent.py   # MedGemma 27B analytics + forecasting
│   │   └── alert_agent.py          # Escalation, FHIR reports, advisories
│   ├── models/
│   │   ├── __init__.py
│   │   ├── medgemma_loader.py      # Model loading utilities (4B, 27B, quantized)
│   │   ├── medasr_loader.py        # MedASR speech-to-text
│   │   ├── medsiglip_loader.py     # MedSigLIP image encoder
│   │   └── forecaster.py           # Time-series forecasting (diffusion / statistical)
│   ├── data/
│   │   ├── __init__.py
│   │   ├── syndromic_schema.py     # Pydantic schemas for extracted signals
│   │   ├── synthetic_generator.py  # Generate training data for LoRA fine-tune
│   │   └── disease_definitions.py  # WHO ICD-10 syndromic surveillance categories
│   ├── utils/
│   │   ├── __init__.py
│   │   ├── anomaly_detection.py    # CUSUM, Poisson thresholds
│   │   ├── fhir_generator.py       # FHIR-compatible report output
│   │   └── geo_utils.py            # Location handling
│   ├── api/
│   │   ├── __init__.py
│   │   └── app.py                  # FastAPI backend
│   └── frontend/                   # React dashboard (later)
├── notebooks/
│   ├── 01_syndromic_extraction_baseline.ipynb
│   ├── 02_lora_finetune_syndromic.ipynb
│   ├── 03_image_triage_demo.ipynb
│   ├── 04_surveillance_analytics.ipynb
│   └── 05_full_pipeline_demo.ipynb
├── scripts/
│   ├── generate_training_data.py
│   ├── run_finetune.py
│   └── run_evaluation.py
├── tests/
│   └── test_extraction.py
└── docs/
    ├── technical_overview.md        # 3-page writeup for submission
    └── architecture_diagram.png

---

Fine-Tuning Strategy

What Gets Fine-Tuned (and Why)

MedGemma 4B → LoRA fine-tune for Syndromic Extraction

MedGemma is trained for clinical Q&A, radiology reporting, and medical knowledge. It is NOT trained to extract structured epidemiological signals from free-text clinical narratives. We need it to:

• Parse "patient presents with 3 days of watery diarrhea, vomiting, and dehydration" into structured JSON:

  {
    "symptoms": ["watery_diarrhea", "vomiting", "dehydration"],
    "syndrome_category": "acute_watery_diarrhea",
    "severity": "moderate",
    "onset_days": 3,
    "age_group": "adult",
    "icd10_codes": ["A09", "R11", "E86"],
    "reportable_conditions_flagged": ["cholera_suspect"]
  }
  

Training Data: We generate ~2,000 synthetic clinical encounter → structured extraction pairs covering WHO's priority syndromic surveillance categories:

• Acute watery diarrhea (cholera signal)
• Acute bloody diarrhea (dysentery)
• Acute febrile illness (malaria, dengue, typhoid)
• Acute respiratory infection (influenza, COVID, pneumonia)
• Acute neurological syndrome (meningitis, encephalitis)
• Acute rash with fever (measles, chickenpox)
• Acute hemorrhagic fever (Ebola, Marburg, dengue hemorrhagic)
• Unexplained cluster of illness

LoRA Config: r=16, alpha=32, targeting all linear layers, QLoRA 4-bit quantization Training: ~1-2 hours on a single H100, using TRL's SFTTrainer

What Stays Frozen

• MedASR: Works out-of-box for medical speech transcription (5.2% WER)
• MedSigLIP: Use as feature extractor with cosine similarity for zero-shot disease image matching
• MedGemma 27B: Prompt-engineered for surveillance reasoning — too large to fine-tune efficiently in 13 days, and prompting is sufficient for report generation

---

Setup & Installation

# Clone and setup
git clone <repo>
cd epicast
pip install -r requirements.txt

# Set Hugging Face token (must have accepted MedGemma terms)
export HF_TOKEN=your_token_here

# Generate synthetic training data
python scripts/generate_training_data.py

# Run LoRA fine-tuning
python scripts/run_finetune.py

# Launch the app
python -m src.api.app

---

Competition Deliverables

1. Video Demo (≤3 min): Walkthrough of a simulated outbreak detection scenario
2. Technical Overview (≤3 pages): Architecture, novel contributions, evaluation results
3. Reproducible Code: This repository + Kaggle notebook

Target Awards

• Main Track: $30K (1st place)
• Novel Task Prize: $5K (syndromic surveillance as a new MedGemma application)

---

License

CC BY 4.0 (as required by competition)