Abstract
Abstract
Conventional assessments of communication and psychological characteristics often rely on standardized questionnaires or manual observations, making objective and scalable evaluation of conversational behavior difficult. This study investigated whether natural language processing (NLP) techniques could quantify characteristics of psychiatrist--participant conversations and provide conversational indicators associated with social communication characteristics and psychological distress. Conversation transcripts collected from psychiatrist-led semi-structured conversations were analyzed using multiple NLP approaches, including lexical similarity, semantic embedding models, lexical richness, conversational statistics, and large language model (LLM)-based extraction of participants' psychological concerns. Associations between conversational features and standardized questionnaire scores were examined using Spearman's rank correlation coefficients with Benjamini--Hochberg false discovery rate correction. Machine learning models were additionally evaluated for predicting both the Social Responsiveness Scale-2 (SRS-2) and the Kessler Psychological Distress Scale (K6). LLM-derived psychological concern agenda features showed relatively stronger associations with the questionnaire measures than conventional lexical overlap measures, while some semantic representations also showed modest associations. In particular, depression-related agenda content showed a significant association with both SRS-2 and K6 scores. The correlation between SRS-2 and K6 scores was ρ = 0.60 (p < 0.001). In machine-learning analyses, conversational features showed modestly better predictive performance for K6 than for the SRS-2. Elastic Net achieved a Spearman correlation of ρ=0.386 between the observed and out-of-sample predicted K6 scores, whereas prediction of SRS-2 scores remained limited across the evaluated models. These findings suggest that conversational NLP features may capture information related to psychological distress more readily than social communication characteristics in the present dataset.