Using Machine Learning to Predict Psychosomatic Rehabilitation Success

Abstract: This study used data from 7,436 patients across seven psychosomatic rehabilitation clinics to explore predictors of treatment success and to assess the feasibility of predicting rehabilitation outcomes using machine learning. Five outcomes based on the biopsychosocial model – activity, depressive symptoms, participation, phobic fears, and somatoform complaints – were derived from the HEALTH-49 and ICF AT-50 Psych. The outcomes were dichotomized to indicate relevant change from admission (T1) to discharge (T2). Random forests using baseline scores, PAREMO-20, and SIMBO-C items, treatment year, clinic, sex, and age, outperformed all dummy classifiers, yielding accuracies between 63.0% (participation) and 75.9% (phobic fears). Feature importances revealed baseline scores as key predictors. Skepticism toward rehabilitation predicted all outcomes except phobic fears, while willingness to change predicted all outcomes except depressive symptoms. Difficulties in social interaction impacted depressive symptoms and phobic fears. Work-related challenges were key predictors for activity and participation. Age, sex, year, and clinic had no impact.