AI-Generated Investigation Guidance for Vaccine Adverse Event Surveillance
A neuro-symbolic AI pipeline that processes passive surveillance reports of suspected vaccine adverse events and generates structured investigation guidance for field epidemiologists.
Interactive Case Report Viewer
4 representative VAERS myocarditis/pericarditis cases demonstrating distinct pipeline pathways:
- Brighton L4 → Early exit with targeted investigation guidance
- WHO A1 → Vaccine-associated with monitoring recommendations
- WHO B2 → Indeterminate with targeted viral testing guidance
- WHO C → Coincidental with escalation for giant cell myocarditis
6-agent, 9-stage neuro-symbolic pipeline:
- Neural perception: LLM extracts clinical observations (Claude Sonnet 4)
- Symbolic reasoning: Deterministic code executes all classification decisions
- Investigation guidance: Structured recommendations via Claude Haiku 4.5
Every classification is reproducible. Every reasoning step is traceable to published evidence.
When data is too incomplete for assessment (20% of cases), the system doesn't guess — it defers. It generates targeted investigation guidance specifying what tests to order, in what priority, and what certainty level each unlocks.
Submitted to JMIR Public Health and Surveillance (2026).
Cheon ME, Son EC. AI-Generated Investigation Guidance for Vaccine Adverse Event Surveillance: Development and Evaluation of a Neuro-Symbolic Causality Assessment Pipeline.
- Pipeline code (MedGemma 4B edge deployment version): github.com/SuahCheon/vax-beacon
- Validation dataset: Kaggle
MIT