Development and External Validation of a Treatment-Adjusted Machine Learning Model for Precision Allocation of Group-Based Depression Care Among People Living with HIV in Uganda.

Background: Group-based depression care in Ugandan HIV services is widely implemented, yet some patients remain symptomatic, and uniform allocation may inefficiently use limited resources. We developed and externally validated a treatment-adjusted machine learning model to predict six-month depression non-remission and inform precision allocation within HIV services.
Methods: We analyzed data from 1,140 adults living with HIV and significant depression symptoms enrolled across 30 HIV clinics in the SEEK-GSP trial (PACTR201608001738234). Participants were assigned to Group Support Psychotherapy (GSP) or Group HIV Education (GHE). The primary outcome was six-month depression non-remission, defined as Self-Reporting Questionnaire (SRQ) score ≥6 and a functional impairment score <9. Treatment assignment was included to enable risk estimation adjusted for treatment. Three machine learning models (Elastic Net, Random Forest, and XGBoost) were trained on baseline data from Gulu and Kitgum and externally validated in Pader district, with the best-performing parsimonious model selected to inform precision allocation. Calibration was evaluated on the external set using slope, intercept, Brier score, and the area under the receiver operating characteristic curve [AUC], with isotonic regression and Platt scaling applied for recalibration.
Results: All models demonstrated excellent discrimination in geographically distinct external validation (AUC 0.974–0.980). XGBoost achieved strong overall classification performance in the Pader cohort (AUC 0.979; accuracy 0.920; sensitivity 0.865; specificity 0.970), supporting reliable identification of individuals at risk of non-remission. Post-calibration isotonic regression improved probability alignment (slope 1.703 to 1.00; Brier 0.070 to 0.015; AUC 0.986). Across modeling approaches, consistent predictors of non-remission included treatment assignment, older age, economic vulnerability (lower income, savings, and employment instability), stigma, low perceived social support, and maladaptive coping patterns.
Conclusion: Treatment-adjusted machine learning models can accurately predict six-month depression non-remission and provide a foundation for precision allocation of group-based depression care within HIV services in low-resource settings.