Using Machine Learning to Predict Treatment Outcome in a Harmonized Dataset of Youth Anxiety Treatments

Machine Learning (ML) is a promising approach to identify predictors/moderators for youth anxiety treatments. To this end, data from nine randomized controlled trials of youth anxiety treatments were harmonized into a dataset (N = 1362; Mage = 10.59, SDage = 2.47; 48.9% female; 71.9% White, 5.9% Black, Other, 5.9%; 10.8% Hispanic) and ML algorithms were used to predict outcomes. Models were then applied on an external validation sample in a research clinic (N = 50; Mage = 12.04, SDage = 3.22; 56% female; 76% Caucasian, 10% Black, 6 % Asian, 2% Other; 6% Hispanic). To examine predictive features by treatment type, Lasso Regression models were built separately for youth who completed individual cognitive behavioral therapy (CBT), family CBT (FCBT), sertraline alone (SRT), and combination of SRT and CBT (COMB). Automatic relevance determination (ARD) emerged as the best performing model in the harmonized (RMSE = 1.84, R2 = 0.28) and external validation datasets (RMSE = 1.87, R2 = 0.11). Predictive features of poorer outcomes were primarily indicators of symptom severity and trial effects, although predictors varied within treatments (e.g., caregiver psychopathology was predictive for FCBT; depressive symptoms were predictive for COMB). Implications for use of ML to identify predictors/moderators are discussed.