A Machine-Learning-Based Matching System For Patient-Therapist Dyads In Routine Psychotherapy

ObjectiveTherapists differ in their effectiveness, yet precision mental health has primarily focused on matching patients to treatments rather than to the person delivering it. We investigated whether machine-learning (ML) models trained on comprehensive patient and therapist data can predict early psychotherapy outcomes and generate patient-specific therapist recommendations. MethodsOur sample included 1,159 patients nested within 189 therapists in routine outpatient psychotherapy in Germany. The outcome was mental distress (PHQ-8 and GAD-7 composite) at week 6 of therapy. Predictors comprised (A) therapist demographics, professional information, and personality, and (B) patient demographics, personality, and clinical features. We trained and compared three explainable ML models with strength in modeling interactions, i.e., Factorization Machines, CatBoost, and Explainable Boosting Machines, against reasonable benchmarks (i.e., patient-only). We evaluated recommendation validity by regressing week-6 distress on the predicted rank of the factual therapist. ResultsA total of 957 patients nested within 174 therapists were available for analysis (5±4 patients per therapist). CatBoost outperformed other ML architectures in predicting mental distress (RMSE = 6.05 ± 0.11) but performed similarly to the patient-only benchmark (RMSE = 6.09 ± 0.12, p = .09). Week-6 distress tended to be higher when the factual therapist received a worse rank, but this association was small and not statistically significant (b = 0.0089, SE = 0.0060, 95% CI [-0.0028, 0.0207], p = .136).ConclusionsTherapist features added little predictive value beyond patient features, and we found no robust evidence that model-based therapist rankings translated into better outcomes. Datasets with higher therapist caseloads are needed to evaluate therapist recommender systems more conclusively.