KAIST and Hospital Deploy LLM Tool to Structure Psychiatric Intake
Researchers at KAIST and Gangnam Severance Hospital have built a conversational AI system to assist initial psychiatric consultations. The tool conducts patient intake interviews before clinicians meet with patients, then generates a clinical summary flagging symptoms and possible diagnoses for clinician review.
The team evaluated the system using 1,440 simulated patient interactions and presented the work at ACM CHI 2026. The approach differs from fixed questionnaire systems by using follow-up questions and counseling techniques-empathy, restatement, and clarification-to draw out patient narratives in real time.
How the system works
The LLM-based tool conducts a conversational intake with patients, cross-referencing their replies against psychiatric knowledge sources. It adjusts its questioning flow in real time based on patient responses, applying counseling-style interaction patterns rather than mechanical yes-or-no formats.
The system produces a clinical dashboard that organizes reported symptoms and candidate disorders. It also generates specific prompts for clinicians-items to verify about sleep, appetite, weight, and other clinical markers during the actual consultation.
What remains unclear
The sources do not disclose the underlying LLM architecture, training data, or safety mechanisms. The team has not published details on how the system handles model hallucination, demographic bias, or the privacy of sensitive health narratives-all known risks in clinical AI applications.
The evaluation used virtual patient dialogues rather than live clinical trials. Real-world validation with actual patients and clinician workflows has not been reported.
What matters for clinical deployment
Whether this research advances to clinical use depends on several factors. Live trials with real patients and integration into clinician workflows would be necessary steps. The team would need to disclose how it grounds the model against reliable medical sources and prevents harmful outputs.
Interoperability with electronic health records and clinician documentation systems would also be required. User-acceptance studies for both patients and clinicians-including performance across demographic groups-would help identify whether the tool works equally well across populations.
Extended peer review and evaluation data beyond the virtual-patient experiment would strengthen the case for deployment.
Context
Conversational AI in clinical intake addresses two operational problems: limited clinician time per patient and inconsistency in how patients report symptoms. Automated intake can improve data completeness and triage consistency, but clinical-AI observers flag concerns around accuracy, bias, privacy, and accountability in clinical settings.
This work represents applied research presented at a major HCI conference. It illustrates industry interest in using generative AI and LLMs to structure subjective clinical narratives, though technical validation and real-world safety controls remain to be demonstrated.
The research is relevant to practitioners working on AI for healthcare, clinical NLP, and clinical AI deployment-but it is not a foundational model release or large-scale deployment announcement.
Your membership also unlocks:
