Identifying CBT non-response among OCD outpatients: A machine-learning approach.

Area of Science:

• Psychiatry
• Clinical Psychology
• Computational Psychiatry

Background:

• Machine learning (ML) models can predict individual treatment outcomes, but utility often depends on data accessibility.
• Few studies utilize low-cost, easily acquired sociodemographic and clinical data for treatment prediction.
• Previous research showed significant but clinically insufficient predictions in a broad diagnostic sample.

Purpose of the Study:

• To evaluate if ML predictions improve in a large, naturalistic, and diagnostically homogeneous obsessive-compulsive disorder (OCD) sample.
• To assess the clinical utility of sociodemographic and clinical data for predicting OCD treatment outcomes.
• To compare prediction accuracy in a homogeneous OCD sample versus a broader diagnostic spectrum.

Main Methods:

• Utilized routinely acquired sociodemographic and clinical data from 533 OCD outpatients undergoing cognitive behavioral therapy (CBT).
• Developed and validated ML models to predict treatment remission and dimensional change.
• Employed balanced accuracy for remission prediction and correlation coefficient (r) for dimensional change.

Main Results:

• The best ML model achieved 65% balanced accuracy in predicting remission on unseen data (p=0.001).
• Higher OCD symptom severity predicted non-remission; earlier age of onset and higher socioeconomic status predicted remission.
• Prediction of dimensional change yielded an r=0.31 (p=0.001) between predicted and actual values.

Conclusions:

• ML predictions in a diagnostically homogeneous OCD sample were not inherently superior to those in a more diverse patient group.
• Future research should explore refined psychological predictors related to disorder etiology and maintenance.
• Integrating additional data modalities like neuroimaging or ecological momentary assessments may enhance prediction accuracy.