Using Machine Learning Algorithms to Predict the Effects of Change Processes in Psychotherapy: Toward Process-Level Treatment Personalization

Abstract

This study aimed to develop and test algorithms to determine the individual relevance of two psychotherapeutic change processes (i.e., mastery and clarification) for outcome prediction. We measured process and outcome variables in a naturalistic outpatient sample treated with an integrative treatment for a variety of diagnoses (n = 608) during the first 10 sessions. We estimated individual within-patient effects of each therapist-evaluated process of change on patient-evaluated subsequent outcomes on a session-bysession basis. Using patients’ baseline characteristics, we trained machine learning algorithms on a randomly selected subsample (n = 407) to predict the effects of patients’ process variables on outcome. We subsequently tested the predictive capacity of the best algorithm for each process on a holdout subsample (n = 201). We found significant within-patient effects of therapist perceived mastery and clarification on subsequent outcome. In the holdout subsample, the best-performing algorithms resulted in significant but small-to-medium correlations between the predicted and observed relevance of therapist perceived mastery (r = .18) and clarification (r = .16). Using the algorithms to create criteria for individual recommendations, in the holdout sample, we identified patients for whom mastery (14%) or clarification (18%) were indicated. In the mastery-indicated group, a greater focus on mastery was moderately associated with better outcome (r = .33, d = .70), while in the clarification-indicated group, the focus was not related to outcome (r = −.05, d = .10). Results support the feasibility of performing individual predictions regarding mastery process relevance that can be useful for therapist feedback and treatment recommendations. However, results will need to be replicated with prospective experimental designs.