Using machine learning to predict long-term treatment outcomes in psychotherapy for personality disorders

Objective: Evidence for predictors of long-term outcomes in personality disorder (PD) treatment remains limited. Machine learning (ML) shows promise in early prediction of treatment outcome to inform clinical practice. This study evaluates the performance of various ML algorithms and examines variables that predict long-term outcomes.Method: Personality functioning was assessed one-year post-treatment in 524 patients who received specialized outpatient psychotherapy for PD, using baseline predictors (demographic and clinical variables at intake) and post-treatment predictors (treatment and clinical variables at treatment end). Regularized linear regression methods were compared with more complex tree-based approaches using cross-validation. Predictive performance was assessed with Mean Absolute Error (MAE) between true and predicted scores.Results: Baseline-only predictors achieved an MAE of 0.40-0.44 (10-11% on a 0-4 scale range). Post-treatment predictors substantially reduced the MAE to 0.27-0.33 (7-8%). Only in the latter case did complex models perform slightly better. Predictors were anhedonia, self-esteem, and relational functioning.Conclusion: Baseline predictors alone provided limited prognostic value, while updating the model with post-treatment predictors improved performance, lending support for a forecasting approach to long-term outcome. In this tabular clinical dataset, complex models provided marginal improvements at best, making simpler models preferable for their interpretability.