Identifying predictors of treatment response in Hoarding Disorder using machine learning

The current gold standard treatment for Hoarding Disorder (HD) is cognitive behavioral therapy (CBT). However, typical symptom reduction is 25% and only about one third of individuals achieve a clinically meaningful response. We used machine learning (ML) to identify predictors of treatment response using data from a non-inferiority trial of clinician vs peer-led group CBT for HD. Assessments were completed for 323 individuals with HD at post-treatment (N = 323) and at one-year follow-up (N = 183). Assessments included neuropsychological testing, symptom severity and psychiatric diagnoses, beliefs about treatment, treatment participation, and adherence measures. Outcomes included changes in HD severity (Saving Inventory-Revised) and functioning (Activities for Daily Living-Hoarding), which were modeled separately. Linear regression and decision tree models were optimized with cross-validation and used to select predictors of response. LASSO was used to distinguish predictors for symptom trajectories (sustained response, regressed response, late response, non-response) determined at one-year follow-up. Optimized models found that reduction in HD severity immediately post-treatment was predicted only by beliefs that treatment was beneficial. Improvement in HD functioning post-treatment was predicted by beneficial beliefs about treatment, memory related saving, and better sustained attention. At follow-up, higher HD severity reduction was predicted by identifying as non-White, more time maintaining progress, and beneficial treatment beliefs. Higher HD functioning reduction at follow-up was predicted by memory related saving, beneficial treatment beliefs and better performance on a sorting task at pre-treatment. Beliefs about treatment, possessions as memory support, sorting, attention, and visual memory may be key factors in treatment response.