Personalized treatment selection in routine care: Integrating machine learning and statistical algorithms to...

Objective: This study aims at developing a treatment selection algorithm using a combination of machine learning and statistical inference to recommend patients' optimal treatment based on their pre-treatment characteristics. Methods: A disorder-heterogeneous, naturalistic sample of N = 1,379 outpatients treated with either cognitive behavioral therapy or psychodynamic therapy was analyzed. Based on a combination of random forest and linear regression, differential treatment response was modeled in the training data (n = 966) to indicate each individual's optimal treatment. A separate holdout dataset (n = 413) was used to evaluate personalized recommendations. Results: The difference in outcomes between patients treated with their optimal vs. non-optimal treatment was significant in the training data, but non-significant in the holdout data (b = -0.043, p = .280). However, for the 50% of patients with the largest predicted benefit of receiving their optimal treatment, the average percentage of change on the BSI in the holdout data was 52.6% for their optimal and 38.4% for their non-optimal treatment (p = .017; d = 0.33 [0.06, 0.61]). Conclusion: A treatment selection algorithm based on a combination of ML and statistical inference might improve treatment outcome for some, but not all outpatients and could support therapists' clinical decision-making.