HealNet

HealNet — MCP Care‑Gap Navigator

Inspiration

I built HealNet because I wanted a small, high-impact demo that proves useful clinical value in under three minutes. I was inspired by real care‑coordination workflows where clinicians spend too much time hunting records; a reliable agent that surfaces prioritized gaps and next steps can save time and reduce missed care.

What it does

HealNet is a focused MCP (Model Context Protocol) superpower that summarizes a patient’s chart, detects actionable care gaps, and drafts concise follow‑up plans for clinicians and care coordinators. I built it to show a realistic, interoperable healthcare workflow that judges and clinicians can immediately understand and trust.

How we built it

Core server: server.py — a FastMCP-based MCP server exposing tools:

summarize_patient_chart
identify_care_gaps_tool
draft_follow_up_plan
generate_patient_message
- FHIR adapter: fhir.py pulls patient data (live FHIR when configured, bundled demo fallback otherwise).
- Care-gap logic: care_gaps.py implements deterministic checks (e.g., stale HbA1c, elevated BP, missing screenings) and returns prioritized gaps with evidence.
- Optional LLM: configured by env vars (HEALNET_LLM_BASE_URL, HEALNET_LLM_API_KEY, HEALNET_LLM_MODEL) and used only to create succinct summaries when available.
- Packaging & docs: pyproject.toml, requirements.txt, README.md, and a demo_bundle.json so the app runs for demo without a real FHIR server.
- Deployment & demo: local run + Cloudflare Quick Tunnel for public endpoint and integration into Prompt Opinion for live recording.

Challenges we ran into

Quota/config friction: free‑tier model quotas (Gemini/GitHub) required me to support multiple provider options and graceful LLM disablement when unavailable.

FHIR variability: different servers and resource shapes forced robust parsing and many defensive checks.
Time & scope: balancing a complete, testable demo with safety (avoid hallucination in clinical outputs) while delivering within hackathon time.

Accomplishments that we're proud of

Deterministic core logic for safety: care gaps include evidence fields and priority levels (HIGH/MEDIUM/LOW).

LLM only for human-friendly summarization; outputs always return structured fields suitable for UI automation or audit.
Demo fallback bundle (demo_bundle.json) ensures reproducible results for judges.

What we learned

How MCP and SHARP-style context propagate patient IDs and FHIR tokens into a tool-first workflow.

Practical tradeoffs between deterministic logic and LLM polishing: use deterministic rules for safety-critical gap detection and an LLM for concise, human-friendly phrasing.
How to make an app resilient: graceful FHIR fallbacks (demo bundle), environment-based LLM configuration, and robust parsing of heterogeneous FHIR resources.
Devops lessons: packaging a tiny Python MCP server, running it locally, exposing it via Cloudflare Quick Tunnel, and wiring it into Prompt Opinion for an in‑platform demo.