This multicenter, prospective study uses machine learning to identify the most prominent multilevel risk factors of problematic substance use (PSU) among a sample of 200 Ugandan youth living with HIV (YLHIV) ages 18-24. The specific model classes applied in this study were logistic regression, random forest, gradient boosting, XGBoost, support vector machine and AdaBoost. Risk factors at the individual (demographics, mental health status), interpersonal, familial (household size), and community or societal levels (promotion of alcohol) were assessed. 10-fold cross validation with 10 repetitions was performed, training each model on 9 of the 10 subsets of the n = 200 dataset. Based on these 9 subsets, each model was tested in assigning a PSU of 0 (indicating no PSU) or 1 (indicating PSU) to the remaining subset. The random forest model was determined to have the best predictive performance (AUROC = 0.78). After evaluating the most effective model class, researchers found that depression, number of sexual partners, and whether the individual's family sells or makes alcohol were the most common across folds, but 27 out of 50 risk factors were never selected in any fold. A key limitation is the lack of external or temporal validation, which restricts the study’s generalizability beyond this specific population and time.