JMIR Publications

ABSTRACT

Background:

Rapid advances in large language models (LLMs) have expanded interest in healthcare applications that require complex information processing and decision support. Digital health assistants and symptom checkers, which have historically been rule-based, are increasingly incorporating AI capabilities to support initial symptom assessment and patient triage. An accurate and reliable AI-enabled triage tool could improve patient navigation, reduce unnecessary health care utilization, and support earlier recognition of clinically serious conditions.

Objective:

The objective of this study was to analytically validate the performance of the Personal Health Assistant (PHA), a Large Language Model (LLM)-based patient support tool in producing appropriate recommendations using simulated patient encounters and expert physician review.

Methods:

We conducted a prospective analytical validation study of the PHA using synthetic patient data. The evaluation set was constructed from 772 synthetic cases generated from published triage protocols and persona-based LLM-assisted generation. Cases included patient vignette summaries with medical histories and simulated patient conversations. PHA provided guidance and recommendations based on these inputs, which were compared to an independent ground-truth derived from expert physician review. Co-primary endpoints of urgent undertriage, nonurgent undertriage, and overtriage were each evaluated against prespecified clinical performance thresholds.

Results:

The final analysis dataset contained 772 synthetic cases. Urgent undertriage was 28/406 (6.9%, 95% CI 4.6%-9.8%), nonurgent undertriage was 56/305 (18.4%, 95% CI 14.2%-23.2%), and overtriage was 40/364 (11.0%, 95% CI 8.0%-14.7%). Overtriage met the prespecified performance threshold (<30%), whereas urgent (<5%) and nonurgent (<15%) undertriage thresholds were not met.

Conclusions:

PHA maintained acceptable overtriage but did not meet prespecified undertriage targets. These findings support the value of structured predeployment analytical validation as an early evidence step for patient-facing AI systems and highlight its utility for iterative refinement; while also underscoring the need for prospective clinical validation in light of the inherent limitations of simulation-based studies.