Evaluating motivational interview quality using large language models and hidden Markov models

Motivational Interviewing (MI) is a counseling approach that promotes behavior change by eliciting "change talk" and minimizing "sustain talk." Traditional methods for assessing MI quality, such as manual coding, are labor-intensive, subjective, and difficult to scale. This study introduces an automated framework integrating large language models (LLMs) and Hidden Markov Models (HMMs) for evaluation of MI session quality.

This study evaluates the effectiveness of an LLM-HMM framework in predicting MI session quality and examines motivational state transitions in high- and low-quality sessions.

A dataset of 40 MI sessions was analyzed. Client utterances were classified and numerically scored by an LLM based on their intention toward or away from change. With HMMs, we used these scores to examine the motivational state transitions across each session. Differences between high- and low-quality sessions were quantified by comparing transition matrices using Frobenius norms. Statistical significance was assessed via a permutation test. Predictive performance was evaluated using logistic regression with leave-one-out cross-validation (LOOCV), where transition matrix elements served as independent variables and interview quality as the dependent variable.

High-quality MI sessions exhibited fluid transitions between motivational states, whereas low-quality sessions showed persistence in resistance-oriented states. A statistically significant difference in transition matrices was observed between session groups (p < 0.001). The framework achieved a mean LOOCV accuracy of 0.80, demonstrating strong predictive performance in identifying MI session quality.

This study presents a scalable, objective alternative to manual MI evaluation. Future applications may include real-time therapist support, training, and prognosis prediction, pending further validation on field-collected data.