JMIR Publications

ABSTRACT

Background:

Passive vaccine safety surveillance systems often generate clinically incomplete adverse event following immunization (AEFI) reports, which may lack the diagnostic evidence needed for causality assessment. While the period for collecting critical clinical data is limited, specialist expertise to identify necessary evidence at the point of reporting is not often available. Currently, no existing system provides the structured guidance to evaluate whether a report contains sufficient evidence for assessment or to identify the specific clinical data required.

Objective:

This study aimed to develop and evaluate a surveillance support system that generates actionable investigation guidance for field epidemiologists at the point of AEFI report intake. the system identifies what clinical evidence is present, what is missing, and what additional data would most impact the causality assessment.

Methods:

We developed Vax-Beacon, a 6-agent neuro-symbolic pipeline that processes Vaccine Adverse Event Reporting System (VAERS). The system utilizes large language model (LLM) for generating free-text narratives through clinical observation, curated knowledge database for differential diagnosis matching, and deterministic code for WHO causality classification, producing structured investigation guidance for each case. We tested the system on 100 purposively curated VAERS myocarditis/pericarditis cases. Two field epidemiologists independently evaluated pipeline-generated guidance using 5-point Likert scales and open-ended feedback.

Results:

The pipeline processed all 100 cases without errors. WHO classification yielded A1 in 45%, C in 27%, B2 in 7%, and Unclassifiable in 21%. Brighton Level 4 early exit occurred in 20% of cases precluding definitive classification. For these cases, the pipeline generated prioritized diagnostic checklists specifying which tests would upgrade certainty. Cardiac biomarkers such as troponin I, CK-MB were recommended as high-priority tests and cardiac magnetic resonance imaging as a lower-priority follow-up for suspected myocarditis. The neuro-symbolic architecture ensured 100% reproducibility of all classification decisions across independent benchmark runs. In structured expert review (two field epidemiologists), Likert scores ranged from 3 to 5 (mean 4.33); both reviewers estimated 30–50% workload reduction and agreed the system is suitable as an official investigation support tool.

Conclusions:

Vax-Beacon demonstrates that neuro-symbolic AI can function not as a classification oracle, but as a surge-ready investigation focus tool — directing field epidemiologists to the right evidence items for known adverse events at the moment when collecting that evidence remains feasible. This principle, Designed Deference, addresses a critical gap in passive surveillance: the loss of retrievable clinical evidence between reporting and expert review. Clinical Trial: .