A Longitudinal Analysis of Multimodal Working Alliance Detection in Therapist-Patient Psychotherapy

Abstract

A central component determining the outcome of psychotherapy is the working alliance between a therapist and their patient. This study explores an approach to understanding and predicting this alliance through automatic linguistic analysis of therapy session transcripts, as opposed to traditional self-report measures such as the working alliance inventory.

The methodology encompasses a pipeline for the automatic transcription, diarization and identification of participants in therapy sessions, followed by affect and sentiment analysis at a speaker-level using a custom-trained Dutch language model, which shows promise in capturing affective trends.

Following this analysis, several features are extracted such as the emotional valence, arousal, sentiment and speaker synchronies, aiming to predict working alliance inventory scores and their sub-components of bond, goal and tasks. Significant correlations between some of these features and the alliance scores are revealed, particularly a patient's average valence.

While the final predictive power of the presented models is lacking, valuable insights are gained into the issues surrounding such automatic analysis and prediction. The contribution of this study to computational psychotherapy research is therefore mainly a proof of concept for language based working alliance evaluation.