Baseline Predictors of Depression Non-Remission Six Months After Group-Based Interventions Among People with HIV in Uganda: A Multi-Model Machine Learning Analysis

Background: Group Support Psychotherapy (GSP) improves depression and HIV outcomes among people living with HIV, yet some participants experience persistent depression. We developed and externally validated prognostic models using baseline sociodemographic, clinical, trauma, and psychosocial factors to predict six-month depression non-remission and identify key risk and protective factors. Methods: We analyzed data from 1,140 people living with HIV (PLWH) with mild-to-moderate depression enrolled in the SEEK-GSP cluster-randomized trial (PACTR201608001738234) across 30 HIV clinics in northern Uganda. The primary outcome was depression non-remission at six months, defined as a Mini International Neuropsychiatric Interview (MINI) score ≥5 or a functional impairment score <9. An Elastic Net regularized logistic regression model was trained on baseline data from two districts (Gulu and Kitgum) and externally validated on a geographically distinct test set from a third district (Pader). To assess robustness and convergence in predictor importance, we conducted sensitivity analyses using Random Forest and XGBoost classifiers. Results: In the Elastic Net model, predictors of six-month depression non-remission were assignment to the Group HIV Education arm, unemployment, low perceived social support, alcohol use, and excessive fatigue, whereas assignment to Group Support Psychotherapy, pre-intervention savings group membership, and trauma-related dissociation (difficulty recalling traumatic events) were associated with reduced risk. Random Forest and XGBoost results were broadly consistent. Three variables—treatment assignment, trauma-related dissociation, and marked hopelessness about the future—were consistently selected across all models. Features selected by at least two models included monthly income, employment status, savings group membership, low perceived social support, alcohol use, increasing age, and excessive anger, reflecting convergence across key clinical, psychosocial, and economic domains. Conclusion: These findings lay the foundation for developing risk stratification tools to inform clinical decision-making and prioritize follow-up for individuals at elevated risk of poor mental health outcomes after group-based interventions in resource-limited HIV care settings.