Advancing AI Trustworthiness Through Patient Simulation: Risk Assessment of Conversational Agents for Antidepressant Selection
This addresses the need for scalable risk assessment of AI in healthcare, particularly for antidepressant selection, though it is incremental in applying simulation to a specific domain.
This paper tackled the problem of evaluating healthcare conversational agents by introducing a patient simulator that generates realistic interactions to assess performance and identify errors, revealing a degradation in AI decision aid accuracy from 47.9% for limited health literacy to 81.6% for proficient health literacy.
Objective: This paper introduces a patient simulator designed to enable scalable, automated evaluation of healthcare conversational agents. The simulator generates realistic, controllable patient interactions that systematically vary across medical, linguistic, and behavioral dimensions, allowing annotators and an independent AI judge to assess agent performance, identify hallucinations and inaccuracies, and characterize risk patterns across diverse patient populations. Methods: The simulator is grounded in the NIST AI Risk Management Framework and integrates three profile components reflecting different dimensions of patient variation: (1) medical profiles constructed from electronic health records in the All of Us Research Program; (2) linguistic profiles modeling variation in health literacy and condition-specific communication patterns; and (3) behavioral profiles representing empirically observed interaction patterns, including cooperation, distraction, and adversarial engagement. We evaluated the simulator's effectiveness in identifying errors in an AI decision aid for antidepressant selection. Results: We generated 500 conversations between the patient simulator and the AI decision aid across systematic combinations of five linguistic and three behavioral profiles. Human annotators assessed 1,787 medical concepts across 100 conversations, achieving high agreement (F1=0.94, \k{appa}=0.73), and the LLM judge achieved comparable agreement with human annotators (F1=0.94, \k{appa}=0.78; paired bootstrap p=0.21). The simulator revealed a monotonic degradation in AI decision aid performance across the health literacy spectrum: rank-one concept retrieval accuracy increased from 47.9% for limited health literacy to 69.1% for functional and 81.6% for proficient.