Building ClearPath: An AI Pre-Op Anesthesia Clearance Agent

The Brief

The Healthcare AI Endgame brief asked for a production-grade clinical agent. Not a demo, not a toy. It had to run on the Prompt Opinion platform, speak A2A v0.3, declare its FHIR scopes, and produce output a clinician could actually use. That set of constraints is what pushed us toward a perioperative use case: clearance is high-frequency, structured, evidence-driven, and the inputs map cleanly to FHIR resources we could pull on demand.

How It Works

The pipeline is four stages. Each one is independently testable, which mattered a lot during the hackathon. When something went wrong, we needed to know whether it was a FHIR parsing bug, a rule-engine bug, an LLM reasoning bug, or a protocol bug.

• 01FHIR ingestion: fetches Patient, Condition, MedicationRequest, Procedure, DocumentReference, Observation, Encounter, and AllergyIntolerance over FHIR R4, with refresh-token support for long-lived sessions
• 02Normalization: collapses the bundle into a PatientSnapshot with classified medications (drug class, specialty mapping), recent vitals and labs, and parsed PCP and specialist notes
• 03Clinical engines: evaluates Tier-1 hard-stop triggers, Tier-2 risk factors, and computes the RCRI cardiac score, then escalates based on the procedure's institutional protocol
• 04Claude Sonnet reasoning: enriches the deterministic engine output with clinical narrative, named-guideline citations, specialist referral logic, and plain-language next steps

The response goes back over A2A as a structured Task artifact: a markdown summary in one part, a typed ClearanceOutput JSON in another. The SHARP/FHIR context extension is propagated back, so any downstream agent in the same Prompt Opinion session inherits the same patient context without re-authenticating.

Deterministic First, LLM Second

The biggest architectural call was where to draw the line between deterministic code and LLM reasoning. We landed on a hard rule: the disposition is decided by deterministic engines, never by Claude. Risk scores, hard-stop triggers, anticoagulation flags, RCRI: all computed in Python against published thresholds. Claude is responsible for narrative, specialist selection nuance, and next-step phrasing, but it cannot change the disposition the engines produced.

That separation is what makes the output defensible. If a reviewer asks why a patient was flagged for a cardiology consult, we can point to a named rule, a specific medication, and a published guideline, not a model's vibe.

Dispositions

• 01NO_CLEARANCE_NEEDED: healthy patient, low-risk procedure, no triggers
• 02CLEARANCE_RECOMMENDED: cleared with PCP-level medical clearance
• 03SPECIALIST_REQUIRED: one or more specialist consults required before surgery
• 04ANESTHESIA_REVIEW_REQUIRED: anesthesia team must review before clearance
• 05INSUFFICIENT_INFORMATION: chart is too sparse or FHIR token expired

Clearance Letters

Beyond the clearance decision itself, ClearPath can draft the actual clearance correspondence: one letter to the PCP for general medical clearance, or one focused letter per specialist when multiple consults are needed. Letters are generated in parallel and personalized with the patient's name, DOB, and the PCP's actual name extracted from FHIR DocumentReference.author. Nothing is invented; if the chart doesn't have a PCP name, the letter says so.

The Stack

• 01Python 3.11 and FastAPI for the A2A endpoint and JSON-RPC handling
• 02Claude Sonnet (claude-sonnet-4-6) via the Anthropic SDK, with streaming for follow-up Q&A and web-search grounding for guideline citations
• 03FHIR R4 client with OAuth refresh support, talking to the Prompt Opinion FHIR endpoint
• 04A2A v0.3 protocol, A2A v1 agent-card spec, and the SHARP/FHIR Context Extension for cross-agent context propagation
• 05Deployed on Render with the agent card published at /.well-known/agent-card.json

The Demo Patients

We engineered six synthetic FHIR patients to cover the disposition spectrum. Sarah Bennett, 34, healthy, in for a wisdom-tooth extraction, should resolve to NO_CLEARANCE_NEEDED in seconds. David Okafor, 68, atrial fibrillation on rivaroxaban, type 2 diabetes, hypertension, in for a colonoscopy, should land on SPECIALIST_REQUIRED with a cardiology consult and an anticoagulation hold protocol. The other four sit at intermediate points: optimized cardiac history, decompensated CHF, OSA without titration, multi-specialty orthopedic. They're the test suite, and they're also the demo.

Guideline Grounding

Every threshold and every recommendation traces back to a named, published standard. The engine doesn't invent numbers and Claude isn't allowed to invent doctor names or citations. The references are ACC/AHA perioperative cardiovascular guidelines, the ASA preoperative evaluation framework, RCRI, STOP-BANG for OSA screening, ARISCAT for postoperative pulmonary complications, and FDA drug labeling for perioperative medication management. Follow-up questions get web-search grounding so Claude can cite current guidance live.

What We Learned

Three things. First, FHIR data in the wild is messier than the spec implies. Note parsing, in particular, ate more time than the entire LLM layer combined. Second, the deterministic-first split paid for itself the moment we had to debug a wrong specialty recommendation; the engine output made it trivial to localize the bug. Third, A2A is a genuinely useful protocol once you accept that the agent card is the contract: declaring scopes, skills, and extension URIs up front forced us into a cleaner design than we'd have written otherwise.

What's Next

ClearPath is a hackathon build, but it's also a real product surface. The roadmap is wider clinical-guideline coverage, more surgical specialties, deeper FHIR data, and making the agent drop-in compatible with any healthcare platform, fine-tunable on their own protocols and patient data so it fits the way their clinicians already work. The agent card lives at clearpath-htiy.onrender.com/.well-known/agent-card.json if you want to point a client at it.