Leveraging large language models to assess therapeutic technique diversity and its associations with the therapeutic alliance and outcome

OBJECTIVE: This study employed a novel methodology that leverages a Large Language Model (LLM) to automatically detect and annotate therapists' interventions in psychotherapy transcripts, to examine therapists' Therapeutic Technique Diversity (TTD) and its associations with the working alliance and client functioning. METHOD: A dataset of 872 transcribed psychotherapy sessions was analyzed. Therapists' interventions were annotated using a transformer-based model. TTD was quantified using Shannon's entropy index to calculate the diversity of the interventions. Multilevel regression models examined TTD changes within sessions, within dyads (between sessions), and between dyads, and its associations with the Working Alliance Inventory (WAI) and the Outcome Rating Scale (ORS). RESULTS: TTD significantly increased both within and between sessions. A higher TTD within-dyad was positively associated with WAI and ORS scores, although the association with ORS did not remain significant after accounting for session number. No between-dyads effects of TTD were found. CONCLUSION: TTD emerged as a dynamic feature of therapist behavior associated with the quality of the therapeutic alliance and treatment outcome. These findings underscore the importance of maintaining a diverse repertoire of interventions and highlight the potential of LLMs to capture granular therapeutic processes at scale.