Keywords: prospective healthcare, machine learning, risk prediction, HIV, BART, XGBOOST, GBM, stacked ensembles
Prospective healthcare has empirically proven to be a highly effective paradigm that is radically reshaping public health, clinical medicine, and healthcare as a domain. Despite its immense benefits, little has been done in utilizing its promises to guide the current WHO HIV viral load informed care model. To achieve the ambitious UNAIDS 90-90-90 targets, timely detection of future virologic failure is critical in preventing potential patient-level treatment failure, adverse clinical trajectories as well as immunological deterioration. Even more so in resource-limited settings commonly characterized by an acute shortage of healthcare workforce. As such, this paper aims to disseminate some novel approaches that can potentially be utilized in HIV clinical settings to anticipate and mitigate future risks of virologic failure before they manifest. A series of different variants of statistical learning models were trained and evaluated using dataset extracted from an EHR serving over 90,000 HIV patients in Kenya. Due to the groundbreaking nature of this work, domain experts hand-picked over 50 clinically-meaningful predictors consisting of over 256 million clinical and non-clinical observations. Cross-validated metrics such as accuracy, sensitivity, and specificity were used to evaluate the performance of each model. In the end, ensemble methods such as BART, XGBoost, and stacked ensemble techniques turned out among the top classifiers which exhibited high predicted sensitivity and specificity. BART and XGBoost particularly correctly identified over 90% of patients with future risk of virologic failure. Evidence from this study suggests that accurate forecasting of HIV-related clinical events such as virologic failure may have positive health implications through timely administration of targetted holistic intervention.