Toward Flexible Psychiatric History-Taking and Visualization: Exploring Clinician Perspectives with Large Language Models

Paper Title

Toward Flexible Psychiatric History-Taking and Visualization: Exploring Clinician Perspectives with Large Language Models

Publication Info

• Topic area: AI-assisted psychiatric history-taking and clinical information visualization.
• Keywords: psychiatric interview, large language models, clinical dashboard, adaptive questioning, empathy, simulated patients, DSM-5, functional impairments, clinician-centered design, mental health AI.

Background and Problem

• Problem / challenge: Initial psychiatric interviews are constrained by limited time and unpredictable patient responses, making it difficult to gather essential diagnostic information efficiently. Existing conversational agents have not fully addressed the unique challenges of psychiatric history-taking, such as the need for flexible, empathic questioning under time constraints.
• Significance: Effective psychiatric history-taking is critical for diagnosis and treatment planning, but inefficiencies can lead to missed information, reduced patient trust, and compromised care quality.
• Motivation and related work: Prior systems have focused on standardizing validated questionnaires or improving diagnostic accuracy in general medical domains. However, these approaches lack the flexibility and clinician-centered design needed for initial psychiatric interviews. This study aims to address these gaps by incorporating clinicians’ perspectives into system design.

Solution

• Proposed approach: A flexible AI interviewer system that dynamically adapts question flow to prioritize essential clinical information while maintaining conversational empathy. The system includes a clinical dashboard for summarizing and visualizing interview data.
• Novelty:
  1. Adaptive interview design that adjusts to patient input and time constraints.
  2. Integration of listening techniques to reduce psychological burden and foster trust.
  3. Clinician-centered evaluation using simulated patients and expert feedback.
  4. A structured clinical dashboard for efficient review and decision-making.
• Procedure and key techniques:
  • The AI interviewer uses a DSM-5-based knowledge base to guide questioning.
  • A four-stage iterative loop (Ask, Evaluate, Check, Plan) structures the conversation.
  • Listening skills (e.g., reflecting emotion, restating, clarifying) are applied to enhance patient comfort.
  • Simulated patients with diverse disorders and conversational styles are used for safe, scalable evaluation.
  • A clinical dashboard summarizes key information, linking it to the dialogue transcript for easy review.

Results

• Concrete findings:
  • The AI interviewer achieved ≥90% completeness for most single-disorder cases within 30 minutes, with lower rates for comorbid cases like MDD+ADHD (78%).
  • PIQSCA scores from clinicians: Process (3.96/5), Techniques (3.84/5), Information for Diagnosis (3.69/5).
• Advantage over baselines:
  • Higher efficiency in gathering essential information within time constraints compared to prior systems.
  • Empathic responses and structured summaries align with clinical needs, enhancing usability.
• Experiments / evaluation:
  • 1,440 simulated dialogues with patients representing eight disorders and six conversational styles.
  • Expert evaluation with 19 clinicians, including psychiatrists and psychologists, using structured reviews and thematic analysis.
• Limitations and future work:
  • Limited geographic and cultural diversity among clinician participants.
  • Lack of real-world patient testing; simulated patients may not fully capture clinical complexity.
  • System currently supports a limited set of disorders; future work should expand to broader psychiatric conditions and real-world deployment.

Summary

This study introduces an AI interviewer system designed to support initial psychiatric interviews by dynamically adapting to patient input and time constraints while maintaining empathic communication. Evaluations with simulated patients and expert clinicians demonstrated the system’s ability to gather essential clinical information efficiently and provide structured summaries for diagnostic decision-making. While the system shows promise as a clinician-support tool, future work should address its limitations by incorporating diverse clinician and patient perspectives, expanding disorder coverage, and testing in real-world clinical settings.