Bayesian Model Averaging for Health Outcome Prediction in HIV-Positive Patients in Rivers State

Standard single-model prediction approaches for health outcomes in HIV-positive patients in Rivers State are vulnerable to model selection uncertainty, which Bayesian model averaging (BMA) addresses by combining predictions across a model space weighted by posterior model probabilities. This study applies Bayesian model averaging to predict two-year viral load suppression and CD4 count recovery outcomes in HIV-positive patients receiving antiretroviral therapy in Rivers State, South South Nigeria. Clinical data for 820 HIV-positive patients enrolled in ART at University of Port Harcourt Teaching Hospital and three secondary facilities were used. Twenty-two candidate predictor variables were specified. BMA was implemented using the BMA package in R with Bayesian Information Criterion approximation for marginal likelihood. Model-averaged predictions were compared to single best-model predictions by cross-validated AUC and Brier score. BMA yielded superior cross-validated AUC for viral suppression prediction (AUC = 0.83) compared to the single best logistic regression model (AUC = 0.79). Brier score improvement was 11 percent. Baseline CD4 count, ART regimen, and treatment adherence score were the most consistently included predictors across models (posterior inclusion probability above 0.90). Age and WHO clinical stage showed high uncertainty with posterior inclusion probabilities of 0.61 and 0.58 respectively. The study introduces BMA as a methodologically superior prediction framework for Nigerian clinical HIV research and recommends its adoption in national HIV programme outcome modelling conducted by the Federal Ministry of Health. Keywords: Bayesian model averaging, HIV outcome prediction, Rivers State, antiretroviral therapy, clinical statistics