Explainable boosting machine classification of HIV serostatus among young adults in Tanzania: analysis of the 2022-2023 Tanzania HIV impact survey

Background Young adults remain disproportionately affected by HIV in sub-Saharan Africa. We evaluated Explainable Boosting Machine (EBM) models for classifying prevalent HIV serostatus among young adults in Tanzania, compared their performance with penalized logistic regression, and examined predictor contributions. Methods We analyzed data from 9,904 participants aged 15–24 years in the nationally representative 2022-2023 Tanzania HIV Impact Survey. Seven survey-weighted EBM specifications and an elastic-net benchmark were developed using a cluster-partitioned 80% training set, with repeated cluster-level cross-validation for probability recalibration and threshold selection. Performance was evaluated on a 20% held-out test set using discrimination, calibration, and classification metrics, with 95% confidence intervals estimated from jackknife replicate weights (Korn–Graubard intervals for classification proportions), conditional on the fitted models. Feature stability was assessed using 100 cluster bootstrap refits. Results Overall, 115 participants (1.16%) were HIV positive. In the held-out test set, which included 22 HIV-positive participants, the full EBM achieved a ROC-AUC of 0.730 (95% CI: 0.625–0.834) and PR-AUC of 0.067 (95% CI: 0.000–0.147). The elastic-net benchmark achieved a higher ROC-AUC of 0.769 (paired difference: −0.040, 95% CI: −0.063 to −0.017). Excluding HIV prevention predictors reduced ROC-AUC (difference: −0.179, 95% CI: −0.252 to −0.106), while excluding other predictor domains had minimal impact. At the training-selected threshold, the full EBM had a sensitivity of 0.932 (95% CI: 0.715–0.997), specificity of 0.359, and PPV of 0.011, classifying 64.3% of participants as positive; its Brier skill score relative to a prevalence-only model was 0.007. In the sensitivity analysis, removing previous HIV testing alone reduced ROC-AUC from 0.730 to 0.588. Socio-demographic, HIV prevention-related, and behavioral predictors accounted for 53.0% (95% bootstrap interval 43.5–65.5%), 38.1% (25.0–48.9%), and 8.8% (4.5–18.7%) of total feature importance, respectively. In the full EBM, previous HIV testing, sex, residence, and age were the most important predictors. Conclusion EBM classified prevalent HIV serostatus with moderate discrimination and low precision and did not outperform penalized logistic regression in this low-prevalence setting. Much of its discrimination depended on previous HIV testing, which may partly reflect post-diagnosis information. The main value of EBM was in interpreting model predictions, but limited HIV-positive observations constrained the stability of feature-specific estimates.