ED Decision Rules MCP

• 

  mcp

Inspiration

Emergency physicians apply decision rules ~30x per shift (HEART, Wells PE, qSOFA, NIHSS). Misapplication drives unnecessary CT/admit and — worse — unsafe discharge. No MCP server exists that lets an AI agent apply these validated rules to a FHIR Bundle.

What it does

• 100 validated ED decision rules exposed as an MCP Superpower
• LLM Reasoner picks which rules apply, runs them in parallel, reconciles conflicts
• Returns a CDS Hooks 1.0 card that Epic/Cerner can render natively
• Refuses to recommend when no rule fits ("physician review required")
• 26-complaint / 9-organ / 3-age-band coverage matrix routes free-text complaints

How we built it

• Engine: 98 additive + 2 non-additive rules (MELD-Na regression, Tokyo Guidelines categorical), pluggable scoring-strategy registry
• FHIR variable extractor with PHI scrubbing before any LLM call
• SHARP envelope helpers propagate contextId / correlationId / bundleSha256
• Property-based testing caught 5 real data defects during development
• 19,600 randomized fuzz evaluations per run, 100% pass

Challenges we ran into

• Distinguishing additive from regression rules — solved with a strategy registry
• Avoiding "fake conflicts" when extractor lacks data — solved with insufficient-data guard
• Score-range gaps in published rule definitions — caught by property tests, fixed

Accomplishments we're proud of

• 100-rule library with deterministic versioning (SHA-256 per rule)
• Baseline comparison: 4/4 disposition correct vs naive LLM 3/4; 4/4 rule cited vs 0/4
• 219 tests, 0 regressions across 18 commits

What we learned

• Rule-data defects are common; property-based testing catches them where curated tests don't
• "AI Factor" is in the reasoner, not the scoring — rules stay deterministic

What's next

• Engine extension for true logistic-regression rules (GRACE with intercept + sigmoid)
• More rules (~30 high-value additions documented as gaps)
• SMART-on-FHIR OAuth2 launch-context handling

Built With

• a2a
• fhir-r4
• hl7-cds-hooks-1.0
• mcp
• pytest
• python-3.11
• pyyaml
• ruff
• smart-on-fhir-(r4.smarthealthit.org)