Machine Learning and Explainable AI for Predicting Survival and Mortality in People Living With HIV on Antiretroviral Therapy: A Secondary Analysis of the ACTG-175 Trial Dataset

Le résumé fourni par la source

Background: Identifying reliable predictors of mortality and survival in people living with HIV (PLWH) is essential for personalized risk stratification and treatment optimization. This study employed machine learning (ML) and explainable artificial intelligence (XAI) to identify key prognostic factors and develop predictive models for overall survival and mortality in a large PLWH cohort. Methods: We analyzed data from 2139 patients from the AIDS Clinical Trials Group Study 175. For time-to-event analysis, we implemented Cox Proportional Hazards, Random Survival Forest, and Cox Elastic-Net models. For mortality prediction, we employed Elastic Net, Random Forest, and Bayesian Additive Regression Trees (BART). Models were trained on 70% of the data and validated on the remaining 30%. SHapley Additive exPlanations (SHAP) analysis was applied to the best-performing models for interpretability. Model performance was evaluated using C-index, AUC-ROC, accuracy, sensitivity, specificity, and calibration metrics with 95% confidence intervals (CI). No external validation was performed. Results: The Cox Proportional Hazards model demonstrated the best survival prediction performance (C-index: 0.758, 95% CI: 0.723-0.784; 1-year AUC: 0.885, 95% CI: 0.849-0.934), while Elastic Net was optimal for mortality prediction (AUC: 0.751, 95% CI: 0.712-0.793; accuracy: 0.793, 95% CI: 0.758-0.825). CD4 count at 20 weeks consistently emerged as the most influential predictor across all models, with mean absolute SHAP values of 1.745 for survival and 0.129 for mortality. Treatment-related variables, including off-treatment status and specific antiretroviral therapy (ART) regimens, were also consistently important predictors. Symptomatic status showed strong negative contributions (81% for survival, 81% for mortality), while non-White race paradoxically demonstrated protective effects (73% positive contributions). BART demonstrated comparable performance (AUC: 0.743, 95% CI: 0.695-0.785) with excellent calibration (calibration slope: 0.895). Conclusion: Machine learning models, particularly Elastic Net and Cox Proportional Hazards, effectively predict mortality and survival in PLWH, with immunological response, treatment-related factors, and clinical presentation emerging as the dominant predictors. These findings support the potential of ML and XAI for personalized risk stratification in HIV care, emphasizing the critical importance of optimizing ART regimens, ensuring treatment adherence, and addressing modifiable risk factors to improve long-term outcomes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.