Title of Dissertation: Using a nationally representative health data repository to improve retention in care among people with HIV in Nigeria: A computational epidemiology approach

Background: Antiretroviral therapy (ART) has significantly improved the health outcomes of people with HIV (PWH). However, treatment interruptions remain a major challenge, leading to poor health outcomes and an increased risk of HIV transmission. This study aimed to identify predictors of first interruption in treatment (IIT) in PWH and determine if machine learning offers an alternative analytic approach to predicting IIT among PWH in Nigeria.

Objectives: i) To determine the predictors for the first IIT defined as returning for a clinic visit 28 days or more post the next scheduled appointment, in HIV patients recorded in the Nigerian National Data Repository. ii) To determine the usability of selected supervised machine learning classification algorithms in HIV patient retention as measured by the prediction accuracy metrics for predicting the first IIT.

Methods: We conducted a retrospective cohort study of over 924,847 PWH in Nigeria who started ART between 2015 and 2021. Data were extracted from the Nigeria National Data Repository (NDR), a comprehensive database of all patients managed under PEPFAR funding in Nigeria. We applied supervised learning classification algorithms (random forest, extreme gradient boosting, decision trees, and logistic regression) models to the data and assessed their ability to predict the first interruption in treatment correctly.

Results: The median age of the study population was 34 years, and 65.5% were women. Overall, 54.1% of patients experienced at least one treatment interruption. The most common predictors of first interruption and time to first treatment interruption were being female (male aOR=0.90, 95% CI: 0.89-0.91), younger age of 25-34years (aHR=0.92, 95% CI: 0.91-0.93 for 35-44years, aHR=0.91, 95% CI: 0.90-0.92 for 45-54years, aHR=0.89, 95% CI: 0.88-0.90 for 55-64years and aHR=.89, 95%CI: 0.86-0.92 for 65+years ), secondary school education (tertiary education aHR=0.92, 95% CI: 0.91-0.94), and being unemployed (aHR=1.04, 95% CI: 1.03-1.06). Patients who were enrolled in the ART program after the COVID-19 global lockdown were less likely to interrupt their treatment (aHR=0.17, 95% CI: 0.17-0.17).

The machine learning models were able to predict the risk of treatment interruption with an accuracy of up to 90% with machine learning logistic regression along with random forest, XG-Boost, and decision tree algorithms outperforming classical logistic regression with time-based variables being the top predictive features.

Conclusions: Our findings suggest that the predictors of the first interruption and the time to the same event are similar and that machine learning has high predictivity for these outcomes. The machine learning model developed in this study, when applied to patient care with electronic medical records, can be used to identify patients who are at high risk of interrupting their treatment and to provide them with targeted interventions to help them stay in care.