OBJECTIVES: This study aimed to apply machine learning algorithms to predict treatment outcomes in depression based on pre-treatment predictors and early symptom changes to uncover whether additional variance could be explained.

METHODS: We investigated cognitive behavioral therapy outcomes across a large and naturalistic dataset. Data came from a Dutch outpatient clinic in 2019 and comprised a total of 1975 depression patients. The sociodemographic profile, pre-treatment predictors, and early symptom change (i.e. symptom improvement at 5^th session) were used to predict treatment outcome. Different regression-type machine learners were applied to the whole dataset. Partial dependence analysis and the Lasso method were performed as a double-check, while a logistic regression served as benchmark.

RESULTS: In predicting treatment outcome, machine learning obtained explained variances ranging between 48.8% and 51.2%, explaining 4.1% more variance than linear regression. The best performing algorithm was the random forest with 51.2% of explained variance in the test dataset. Early symptom-change and baseline symptom score were the only significant predictors. Adding early symptom change increased explained variance in all models, with the increase ranging from 22.0% to 23.3%. A sensitivity analysis demonstrated that an imputed and balanced dataset failed to further improve model performances.

CONCLUSIONS: Early symptom-change can predict later outcome and to pre-treatment predictors. However, model performance is far from clinical relevance: the best learner could only predict 51.2% of variance. In comparison with linear regression, more sophisticated preprocessing and learning methods did not substantially improve performance. Thus, the available routinely collected data had little predictive power and richer data would be needed, for instance with information on stressful life events, somatic comorbidities or from imaging procedures.