Objectives: Machine learning models predicting treatment outcomes for individual patients may yield high clinical utility. However, few studies tested the utility of easy to acquire and low-cost sociodemographic and clinical data. In previous work, we reported significant predictions still insufficient for immediate clinical use in a sample with broad diagnostic spectrum. We here examined whether predictions will improve in a diagnostically more homogeneous yet large and naturalistic obsessive-compulsive disorder (OCD) sample. Methods: We used sociodemographic and clinical data routinely acquired during CBT treatment of n = 533 OCD subjects in a specialized outpatient clinic. Results: Remission was predicted with 65% (p = 0.001) balanced accuracy on unseen data for the best model. Higher OCD symptom severity predicted non-remission, while higher age of onset of first OCD symptoms and higher socioeconomic status predicted remission. For dimensional change, prediction achieved r = 0.31 (p = 0.001) between predicted and actual values. Conclusions: The comparison with our previous work suggests that predictions within a diagnostically homogeneous sample, here OCD, are not per se superior to a more diverse sample including several diagnostic groups. Using refined psychological predictors associated with disorder etiology and maintenance or adding further data modalities as neuroimaging or ecological momentary assessments are promising in order to further increase prediction accuracy.
Keywords: cognitive behavioral therapy; machine learning; obsessive-compulsive disorder; outcome; random forest; single-case prediction.