Development and External Validation of a Treatment-Adjusted Machine Learning Model to Support Risk-Informed Group-Based Depression Care for People Living with HIV in Uganda

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background Group-based depression care is widely used in HIV services in Uganda, yet some patients remain symptomatic following treatment. We developed and externally validated a treatment-adjusted machine learning model to support risk-informed group-based depression care for people living with HIV (PLWH). 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 received either 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. 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 data. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, calibration slope, intercept, Brier score and decision curve analysis. Sensitivity analyses excluding treatment assignment were conducted to assess the predictive value of baseline characteristics alone. Results Treatment-adjusted models consistently outperformed treatment-excluded models. In external validation, the parsimonious XGBoost model showed the best overall performance (AUC 0.947), compared with Elastic Net (0.924) and Random Forest (0.918), and demonstrated clinical net benefit across relevant decision thresholds. Following Platt-recalibration, XGBoost showed the best overall performance, preserving strong discrimination (AUC 0.940) while improving the Brier score from 0.174 to 0.113 and the calibration slope from 7.533 to 1.093. Treatment assignment was the dominant predictor of depression non-remission risk, while age, HIV-related stigma, acceptance coping, socioeconomic vulnerability, low social support, and trauma-related symptoms also contributed to prediction. Conclusion The externally validated, Platt-recalibrated parsimonious treatment-adjusted XGBoost model provides a promising approach to identifying people living with HIV at increased risk of depression non-remission and informing enhanced care following group-based depression treatment.