Using Machine Learning to Predict Depression Among Adolescents Living with HIV (ALWHIV) in Uganda

Background and purpose: Adolescents living with HIV (ALWHIV) face heightened risk of depression. In Uganda, limited access to mental health services exacerbates depression necessitating innovative solutions for early detection and intervention. Machine learning (ML) provides a promising approach for identifying at-risk adolescents hence improving mental health outcomes. This study aims to use the state-of-the-art ML techniques to predict depression among ALWHIV in Uganda and support early intervention strategies.

Methods: This study analyzed baseline data from 833 ALWHIV aged 10–17 years in Southern Uganda. Data was collected from April 2022 to November 2023. Depression was assessed using a short version of Children Depression Inventory (CDI) and dichotomized to depressed (CDI score ≥3) and non-depressed (CDI score <3). Predictors were selected guided by Socioecological model at individual, interpersonal, and community levels. Seven ML models, including Random Forest, Logistic Regression, Gradient boosting, decision tree, XGBoost, Least Absolute Shrinkage and Selection Operator (LASSO), and Support Vector Machine were evaluated using stratified 10-fold cross-validation. Model performance was measured using Area under receiver operating characteristic curve (AUROC), Area under precision-recall curve (AUPRC), accuracy, sensitivity, specificity, and F1-score. We fixed the sensitivity to 0.80 during model optimization to prioritize the identification of adolescents at risk of depression to ensure high-risk individuals are accurately identified for potential interventions. SHapley Additive exPlanations (SHAP) analysis was used to interpret feature importance.

Results: The mean age was 13.19 years, and majority of the participants were females (55.5%). The depression rate was 30.97%. Among the models, Random Forest (RF) achieved the best performance (AUROC = 0.79, AUPRC= 0.66). At 0.80 sensitivity threshold, RF obtained specificity of 0.66; accuracy of 0.71; precision of 0.52, and F1-score of 0.64. Key predictors of depression included hopelessness, HIV stigma, HIV shame, self-esteem, teacher support, and caregiver support. SHAP analysis emphasized the dominant role of psychosocial factors in predicting depression.

Conclusion and implications: The study findings indicate the effectiveness of machine learning models, particularly Random Forest, in predicting depression among ALWHIV in Uganda. By analyzing psychosocial factors such as hopelessness, HIV stigma, self-esteem, and caregiver support, the model identifies at-risk individuals, facilitating timely early interventions. The findings suggest that integrating machine learning into adolescent HIV care can significantly enhance mental health outcomes in resource-limited settings. Furthermore, this approach informs future policy decisions and research directions aimed at improving mental health care for ALWHIV.