Using machine learning to predict poor adherence to antiretroviral therapy among adolescents with HIV in low resource settings.

Objectives: Achieving optimal adherence to antiretroviral therapy (ART) and viral suppression is still insufficient for attaining the UNAIDS 95-95-95 target of 2030, especially among adolescents with HIV (AWHIV). This study sought to develop a model to predict poor adherence risk among AWHIV and identify associated risk factors.

Design: We utilized machine learning to predict future ART adherence among AWHIV leveraging its ability to analyze complex, multidimensional data.

Methods: We leveraged a dataset from a 6-year (2012-2018) longitudinal randomized control trial (RCT) with 635 AWHIV in Uganda. We evaluated six machine learning models and retained one with the highest area under receiver operating characteristic (AUROC), and area under precision-recall curve (AUPRC). We further identified principal factors associated with ART adherence based on the best model.

Results: The random forest model outperformed others, with mean AUROC: 0.71 [BC 95% confidence interval (CI) (0.69-0.72)] and AUPRC: 0.55 (BC 95% CI 0.53-0.58). The principal risk factors of poor adherence were poor adherence history; poverty; biological relationship to caregiver; self-concept; savings confidence; duration on ART; frequency discussing sensitive topics with caregivers; household size; economic group assignment; and school enrollment.

Conclusion: Our findings support potential use of machine learning methods and sociobehavioral data for predicting poor ART adherence risk among AWHIV. The predictive tool can help identify AWHIV at the highest risk of treatment failure, and enable early targeted interventions. However, the tool is still preliminary and its accuracy could be improved by incorporating HIV phenotypic and clinical data.

Clinical Trial Number: ClinicalTrials.gov ID:NCT01790373.