Machine learning risk stratification to identify people living with HIV at high risk of delayed ART and advanced immunosuppression

Background: Late presentation and delayed antiretroviral therapy (ART) initiation among people living with HIV (PLWH) are associated with increased morbidity, mortality, and transmission risk. Identifying individuals at high risk of delayed care or severe immunosuppression (CD4 < 200 cells/μL) could enable targeted interventions. Objective: To develop and validate an artificial intelligence (AI)-based risk stratification tool using routinely collected data to identify PLWH at high risk of delayed ART initiation or advanced immunosuppression. Methods: We conducted a retrospective cohort study using de-identified data from HIV-positive individuals (2016 onward). A composite high-risk outcome was defined as delayed ART initiation or baseline CD4 < 200 cells/μL. Multiple classifiers-including logistic regression, XGBoost, LightGBM, CatBoost, multilayer perceptron, and HistGradientBoosting-were trained with five-fold stratified cross-validation. A stacking ensemble incorporating six base models and a LightGBM meta-learner was evaluated. Discrimination (AUC-ROC, AUC-PR), calibration, and clinical utility (decision curve analysis) were assessed. Results: Among 5,436 eligible participants, 61.2% were high-risk. The stacking ensemble achieved an AUC-ROC of 0.915, AUC-PR of 0.950, accuracy of 0.815, and F1 score of 0.858. At ≥90% sensitivity, precision was 0.810. Calibration was good across deciles; decision curve analysis showed net benefit superior to treat-all or treat-none strategies. Baseline CD4 count, age, and transmission route were the strongest predictors. Conclusion: An AI-based risk stratification tool using routinely collected data can identify PLWH at high risk of delayed ART or advanced immunosuppression with strong performance and clinical utility. Prospective validation and equity assessment are needed before implementation

Clinical decision support system · Cohort · Cohort study · Health care · Logistic regression · Prospective cohort study · Receiver operating characteristic · Risk assessment · Risk Stratification · HIV Research and Treatment · HIV-related health complications and treatments · HIV/AIDS Research and Interventions

• Stacked generalization

  Open Access•David H Wolpert•Neural Networks•1992

  Add to list

• Machine learning with routine electronic medical record data to identify people at high risk of disengagement from HIV care in Tanzania

  Open Access•Carolyn A Fahey, Linqing Wei et al.•PLOS Global Public Health•2022

  Add to list

• Modeling the Impact of Social Determinants of Health on HIV

  Open Access•Joseph W Hogan, Noya Galai et al.•AIDS and Behavior•2021

  Add to list