Care Gap Agent

An AI agent that detects preventive care gaps from patient FHIR records missing vaccines, overdue labs, and chronic disease monitoring via A2A protocol on Prompt Opinion.

Comment

Inspiration

Healthcare systems lose billions annually to preventable gaps in care — missed vaccines, unmonitored chronic diseases, overdue screenings. Clinicians are overwhelmed and don't have time to manually audit every patient's record. I wanted to build an agent that does this automatically, in real-time, using open standards.

What it does

Care Gap Agent connects to a patient's FHIR R4 record and runs a full preventive care assessment across four domains:

  • Diabetes — checks HbA1c recency and control (ADA Standards of Care 2024)
  • Hypertension — checks BP monitoring and control (JNC 8 Guidelines)
  • Vaccinations — checks influenza and pneumococcal vaccine status (CDC ACIP)
  • Cancer Screening — checks mammogram and colonoscopy based on age/gender (USPSTF 2024)

It returns a prioritized, structured list of gaps with severity, recommended actions, and guideline citations — ready for any downstream agent to act on.

How we built it

  • Google ADK (TypeScript) for the agent runtime
  • A2A protocol for interoperable agent communication
  • FHIR R4 for patient data access (conditions, observations, immunizations, procedures)
  • Prompt Opinion as the multi-agent workspace
  • Built on the official po-adk-typescript starter repo, extending the healthcare agent with a custom rule engine in healthcare_agent/tools/careGaps.ts
  • Deployed locally and exposed via ngrok, connected to Prompt Opinion as an external A2A agent

Challenges we ran into

  • FHIR data is inconsistently structured across servers — we had to handle both SNOMED codes and text matching for condition detection
  • The A2A message schema required a messageId field not shown in some examples
  • Mapping LOINC codes for HbA1c and blood pressure components required careful handling of nested component arrays

Accomplishments that we're proud of

  • Built a fully functional external A2A agent from scratch in one session
  • Real FHIR R4 integration tested against the SMART Health IT public sandbox
  • Clinically accurate rule engine — correctly skips rules that don't apply to the patient's age, gender, and conditions
  • Successfully published to the Prompt Opinion Marketplace

What we learned

  • How FHIR credentials flow through A2A message metadata into session state
  • How to design agent tools that are composable and callable independently or as a full assessment
  • The importance of graceful degradation when FHIR data is unavailable

What's next for Care Gap Agent

  • Add SDOH (Social Determinants of Health) gap detection from Z-codes
  • Deploy to Google Cloud Run for permanent availability
  • Add LDL/statin gap for cardiovascular risk
  • Support population-level gap reporting across patient cohorts ```

Built With

Share this project:

Updates