Developing a Naïve Bayes risk classification machine learning algorithm to predict high viral load in a low-resource setting

Abstract

Access to routine viral load (VL) testing for people living with HIV (PLHIV) remains limited in many low-resource settings. There is a need for pragmatic, data-driven tools that can proactively identify individuals at increased risk of virological failure. This study aimed to identify predictors of high viral load among PLHIV on antiretroviral therapy (ART) and to evaluate the performance of a Naïve Bayes classification algorithm using routinely collected clinical data. We conducted a retrospective case-control study using secondary clinical data from two public ART facilities in Makonde District, Zimbabwe. A total of 530 participants (156 cases with VL ≥ 1000 copies/mL and 374 controls) were included. Logistic regression was used to identify independent predictors of high VL. A Naïve Bayes classification model was developed using significant and clinically relevant predictors. Model performance was evaluated on the development dataset, and sensitivity, specificity, predictive values and overall accuracy were calculated. Independent predictors of high VL included being a child (adjusted OR 6.11, 95% CI: 2.12-9.31), adolescent/young adult (AOR 4.69, 95% CI: 1.83-7.54), single/non-partnered marital status (OR 2.01, 95% CI: 1.02-3.04), nondisclosure of HIV status (OR 2.56, 95% CI: 1.48-4.17), ambulatory functional status (OR 3.09, 95% CI 1.64-5.43), recent weight loss on two consecutive visits (OR 4.48, 95% CI: 2.11-8.03, p < 0.001), and ART duration <5 years (OR 1.78, 95% CI 1.15-2.63). The Naïve Bayes model demonstrated a sensitivity of 78.2% and specificity of 96.5%, indicating that the model may miss approximately one in five individuals (22%) with true high viral load, misclassifying them as low risk. The developed Naïve Bayes algorithm shows potential as a risk classification tool to support prioritization of VL testing in resource-limited settings. However, performance estimates represent apparent (development-only) performance and require validation in independent datasets before implementation.