Patient Self-reports for Explainable Machine Learning Predictions of Risks to Psychotherapy Outcomes

Abstract

Prioritizing the right patients and providing personalized treatment in a timely manner is crucial to improve access to healthcare. In psychotherapy, at least 1 in 3 patients drop out of treatment, with therapeutic alliance among the common predictors. Recommendations to safeguard retention include strengthening the patient-therapist bond through developing shared goals and checking in on progress and treatment path. Using a sample of 11,095 mental health patients from the USA, we used machine learning to develop a clinical support tool for treatment personalization. A gradient-boosted decision tree was trained on patient-reported data to establish global and individual predictions/predictors for early treatment dropout, treatment length, and symptom outcomes conditional on different treatment lengths in out-of-sample patients. The models demonstrated marginal to moderate improvements in performance versus baseline predictions. The resulting decision support tool could assist in the collaborative selection of treatment goals, appropriate treatment intensity, and optimal allocation of resources. Results are discussed in the context of explainable AI emphasizing interpretability in a clinical context.