HealthPost

LangGraph Pipeline

HealthPost orchestrates the patient visit as a LangGraph StateGraph with conditional routing. Each node is a pure function over a typed state dictionary. Conditional edges skip unnecessary steps (e.g., image analysis when no photos are provided), reducing latency.

intake ─→ analyze_images ─→ extract_meds ─→ diagnose ─→ check_drugs ─→ find_alternatives ─→ assess_safety
             (skip if no                                    (skip if no
              images)                                        interactions)

Edge Deployment

• 4-bit NF4 quantization — MedGemma fits in ~4 GB VRAM via BitsAndBytes
• Runs on any CUDA GPU — T4 (HuggingFace Spaces) or consumer GPUs locally
• No cloud dependency — model inference runs entirely locally after initial download
• Mobile-friendly — Gradio UI renders on mobile browsers, no app install required

Safety Features

• Confidence scoring — low-confidence diagnoses trigger automatic referral recommendations
• Drug interaction alerts — severe interactions flagged with alternatives via DDInter
• Differential diagnosis — multiple candidate conditions listed to prevent anchoring bias
• Structured output — every assessment is a validated JSON object (Pydantic schema)
• Red-flag detection — emergency conditions (sepsis, respiratory distress) trigger immediate referral

Why HealthPost?

• One-click workflow — all modules run automatically in sequence
• Offline-ready — designed for low-connectivity health posts
• CHW-friendly — plain-language output, no medical jargon
• Safety-first — drug interactions checked before any prescription

Important Disclaimer

This tool is a decision-support system for Community Health Workers. It does not replace professional medical judgment. All AI-generated recommendations should be reviewed by a qualified clinician.

Built for the MedGemma Impact Challenge 2025