Embodied movement markers of alliance ruptures in psychotherapy: A machine learning study.

Abstract

Objective: Alliance ruptures are central to psychotherapy process and outcome, yet therapists often fail to detect them in real time, particularly withdrawal ruptures, which tend to manifest subtly. This study examined whether machine learning applied to moment-to-moment movement dynamics during psychotherapy sessions can identify markers of alliance ruptures.

Method: We analyzed 474 videotaped psychotherapy sessions from 86 dyads drawn from a randomized controlled trial for major depressive disorder. Alliance ruptures were coded using the Rupture Resolution Rating System (3RS). Patient and therapist movement was quantified using motion energy analysis across head and body regions. Two feature-engineering approaches captured temporal and spatial movement patterns associated with rupture occurrence.

Results: Models showed greater discriminative performance for withdrawal ruptures (mean AUC = .73) than for confrontational ruptures (mean AUC = .56). Therapist head movements were the most robust individual marker across rupture types, whereas detecting withdrawal ruptures required integrating movement information across multiple body regions and both dyad members.

Conclusion: Withdrawal ruptures leave consistent embodied signatures despite being difficult to identify in vivo. These findings advance psychotherapy process research and highlight the potential of machine learning-based movement analysis to identify behavioral and interpretable markers for alliance-focused training and supervision.