A multilevel machine learning algorithm to predict session-by-session outcome for patients receiving cognitive-behavioural therapy

Aims

New innovations in predictive models, such as machine learning, could enhance the effectiveness of measurement-based care systems by generating more accurate session-by-session psychotherapy outcome predictions. In this study, we developed a tree-based model that integrates the strengths of multilevel and machine learning models to predict patients’ trajectories of clinical improvement during cognitive-behavioural therapy (CBT).

Methods

We used a sample of 1008 outpatients who were treated at a CBT university clinic in Germany. The total sample was randomly divided into a training (2/3 of the sample) and a test (remaining 1/3) set. Grounded on patient demographic and clinical information at baseline, we developed a generalized linear mixed model tree algorithm to predict patients' session-by-session outcome change during the first ten sessions. Results: The best-fitting model in the training set identified 10 groups of patients based on their presenting characteristics and improvement trajectories. In the test set, the algorithm resulted in a correlation of 0.65 between the observed and predicted values for the outcome variable (cross-validation R² = 0.42). Developing failure boundaries based on the tree-based approach allowed us to correctly identify 67.6 % of the test set patients who did not reliably improve within the first 15 sessions of treatment. Discussion: This study provides preliminary support for the integration of multilevel and machine learning models via generalized linear mixed model trees. The algorithms developed could help support routine implementation of precision mental health care strategies by informing therapists’ treatment planning and session-by-session responsiveness for different patient subgroups.