ChileWait — Intelligent Oncology Referral Agent

AI agent that recommends optimal cross-regional referrals for Chile's 2.5M patients on public health waitlists, using FHIR data and RNLE synthetic records.

Comment

Inspiration

Chile has 2.5 million patients waiting for specialist care in the public health system. In 2024, 36,262 people died while on a waitlist. The median wait time for an oncology consultation is 255 days — but this average hides dramatic regional inequality. A patient in Araucanía may wait 389 days while a slot in Maule is available in 52 days. No existing system detects this gap or recommends the transfer.

What it does

ChileWait is an MCP server that exposes four AI-powered healthcare tools to the Prompt Opinion platform:

  1. get_patient_profile() — retrieves a synthetic FHIR R4 patient record with oncology diagnosis, coverage type (FONASA/GES), and clinical urgency score
  2. query_waitlist() — queries synthetic RNLE data for current wait times across all 16 Chilean health service regions
  3. recommend_referral() — the core AI tool: crosses patient profile + regional capacity + coverage constraints to generate a ranked referral recommendation with plain-language clinical justification
  4. generate_heatmap_data() — produces structured data for a regional heat map showing waitlist pressure across Chile (RED/AMBER/GREEN by region)

When a clinician asks "What are this patient's referral options?", ChileWait invokes all four tools and returns: top 3 regional recommendations, days saved per option, coverage compatibility, and a national heat map of oncology waitlist pressure.

How we built it

  • MCP Server: Python + FastMCP framework, deployed on OVHcloud VPS (Ubuntu 22.04)
  • FHIR Data: Synthetic patient profiles generated with Synthea, calibrated to Chilean oncology epidemiology (DEIS-MINSAL 2002-2022)
  • Waitlist Data: Synthetic RNLE dataset based on published MINSAL statistics — 16 regions × 8 pathology groups × 4 coverage types
  • Platform: Integrated and published to Prompt Opinion Marketplace for discovery and invocation

Challenges we ran into

The main challenge was building a recommendation engine that balances three competing factors simultaneously: clinical urgency (how critical is the patient's condition), wait time savings (how much time is saved by transferring), and logistical feasibility (distance, coverage compatibility). Traditional rule-based systems cannot weigh these contextually — this is where generative AI reasoning is essential.

Accomplishments that we're proud of

  • First AI agent specifically designed for Chile's public health waitlist problem
  • Complete FHIR R4 compliance using only synthetic data — zero PHI
  • Covers all 16 Chilean health service regions
  • Directly addresses a gap confirmed by MINSAL's own 2024 procurement for AI-assisted waitlist management

What we learned

Building healthcare AI for a Latin American public system requires understanding the specific coverage structure (GES vs Non-GES, FONASA tiers A-D) and the regulatory context. The solution must be explainable — clinicians need to understand why a region is recommended, not just which region.

What's next for ChileWait — Intelligent Oncology Referral Agent

Phase 2: Pilot with one regional health service using anonymized live RNLE data. Phase 3: Expand to all 29 surgical specialties. Phase 4: Adapt for other Latin American public health systems (Colombia SISPRO, Peru FISSAL).

Built With

  • fastapi
  • fastmcp
  • fhir-r4
  • mcp-protocol
  • minsal
  • ovhcloud
  • prompt-opinion
  • python
  • rnle
  • sqlite
  • synthea
Share this project:

Updates