Optimal Question Selection from a Large Question Bank for Clinical Field Recovery in Conversational Psychiatric Intake

Researchers from the University of Maryland and collaborators have published a paper on arXiv (2604.22067) that tackles a critical challenge in conversational AI for healthcare: how to optimally select questions during psychiatric intake interviews. They formulated this as a question-selection problem using a bank of 655 clinician-authored questions and created synthetic patient vignettes simulating five different behavioral conditions (e.g., guarded, concise). In their evaluation across 300 interview sessions, they compared three strategies: random questioning, a clinically ordered fixed form (mimicking standard intake forms), and an LLM-guided adaptive policy.

The results showed that the clinically ordered fixed form substantially outperformed random questioning, but the LLM-guided adaptive policy achieved the strongest overall recovery of target clinical information. Notably, the advantage of adaptive questioning grew sharply under patient behaviors less amenable to field recovery, especially in guarded-concise conditions. This suggests that performance in clinical conversational systems depends not only on language understanding after information is disclosed but critically on whether the system reaches the right topics within a limited interaction budget. The benchmark provides a controlled framework for studying how clinical structure and adaptive follow-up contribute to information recovery in interactive clinical machine learning.