Prediction of loss to follow-up in postpartum others living with HIV

Fielding, Christopher

Prediction of loss to follow-up in postpartum others living with HIV

Thesis / Dissertation

2025

Publisher

University of Cape Town

Department

Department of Public Health and Family Medicine

Faculty

Faculty of Health Sciences

Abstract

Introduction: In South Africa, postpartum women living with HIV are at an increased risk of loss to follow-up, which poses a critical barrier in vertical transmission prevention. Previous studies have focused on the aetiology of loss to follow-up, with little emphasis on predictive methods. The aim of this study was to use machine learning methods, applied to routine care data, to predict loss to follow-up among postpartum women living with HIV. Methods: This study is a secondary data analysis of 333 peripartum women living with HIV enrolled in the Routine Electronic Mother-Infant Data study in Gugulethu, Cape Town. Data from routine medical records obtained in the parent study were analysed using descriptive statistics, and several machine learning models.. An extreme gradient boosting model was developed and validated to predict the risk of loss to follow-up within the first 9 months postpartum based on routinely available patient data at the point of discharge after delivery. Model calibration was performed on the trained extreme gradient boosting model (n=233), and calibration performance validated on the validation dataset (n=100). Sensitivity and specificity trade-offs were examined and the Youden Index used to identify the optimal classification threshold (i.e., threshold that maximised sensitivity and specificity). Results: Key factors associated with being lost to follow-up included younger maternal age, shorter duration from HIV diagnosis to antiretroviral therapy initiation, and not actively being on antiretroviral therapy at estimated conception. The extreme gradient boosting model demonstrated an area under the receiver operating characteristic curve of 0.721 when validated on the validation dataset, indicating good predictive performance. Model calibration did not significantly improve when calibration methods were applied. Youden Index calculations indicated that the optimal classification threshold was 0.252, providing a sensitivity of 0.827 and a specificity of 0.634. Conclusions: This study emphasises the importance of antiretroviral associated behavioural and healthcare factors in predicting loss to follow-up among postpartum women living with HIV. The developed predictive model showed good predictive power and could assist healthcare providers in identifying high-risk individuals, allowing targeted preventative measures that cost-effectively improve vertical transmission prevention. Future research should focus on validating this model in larger and more diverse populations and integrating it into existing healthcare practices.