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

Area of Science:

• Public Health
• Infectious Diseases
• Biostatistics

Background:

• Routine viral load (VL) testing for people living with HIV (PLHIV) is limited in low-resource settings.
• Predictive tools are needed to identify individuals at risk of virological failure.
• This study addresses the need for data-driven approaches to manage HIV treatment.

Purpose of the Study:

• Identify predictors of high viral load (VL) among PLHIV on antiretroviral therapy (ART).
• Evaluate the performance of a Naïve Bayes classification algorithm for risk stratification.
• Support prioritization of VL testing in resource-limited settings.

Main Methods:

• Retrospective case-control study using secondary clinical data from Zimbabwe.
• Logistic regression to identify independent predictors of high VL.
• Naïve Bayes classification model developed using significant predictors.

Main Results:

• Key predictors of high VL included being a child or adolescent, single marital status, non-disclosure of HIV status, ambulatory functional status, recent weight loss, and ART duration <5 years.
• The Naïve Bayes model achieved 78.2% sensitivity and 96.5% specificity.
• The model may misclassify approximately 22% of individuals with high VL.

Conclusions:

• The Naïve Bayes algorithm shows potential as a risk classification tool for prioritizing VL testing.
• The tool can aid resource-limited settings in managing HIV treatment effectively.
• Further validation in independent datasets is required before implementation.