Explainable machine learning for viral rebound prediction in patients on antiretroviral therapy

Abstract

HIV continues to be a global public health concern, as shown by the approximately 2.1 million HIV-positive patients who were receiving ART without experiencing viral suppression in 2022, according to UNAIDS. Although ART has significantly reduced HIV-related morbidity and mortality, viral rebound (VR) continues to threaten treatment success. In low-resource settings like Uganda, clinical systems have limited data to identify high-risk patients before VR occurs, which increases the likelihood of HIV transmission, treatment failure, drug resistance, and other challenges. Examining VR feature interaction patterns by healthcare workers is challenging due to complex interplay of factors associated with VR. Previous works on VR contributing factors have used less interpretable methods that cannot be trusted and adopted by healthcare workers. Demographic, clinical and laboratory data was preprocessed and divided into chunks of 70%, 15%, and 15% in order to train, validate, and test the six ML models (Hard Voting, Gradient Boosting, LightGBM, Random Forest, XGBoost, and Stacking), respectively. To address class imbalance, SMOTE was applied to the training data, before ML models were trained. Although Hard Voting achieved the highest F1-score (51.56), LightGBM (F1-score of 50.83) was selected as the best-performing ML model due to its superior interpretability when paired with XAI techniques (SHAP, LIME, ELI5, and PDP). Key predictors included past two or more consecutive suppressions, regimen history, adherence patterns, and ART duration. This study demonstrates that ML can provide accurate predictions of VR in ART patients, that can aid in the efficient allocation of healthcare resources by identifying which patients are most at-risk. VR feature interaction patterns will aid in creation of a valuable clinical decision support tool that will notify healthcare workers of patients that need tailored therapies and devise measures to promote viral suppression. XAI techniques integrated ensure that medical professionals understand prediction results and foster proper trust, ultimately strengthening HIV care.