Predictive modelling to identify patients at high risk of virological failure in Kenya

Abstract

Virological failure (VF) remains a significant challenge in HIV treatment, necessitating the development of accurate risk prediction models. Traditional models have faced criticism for their limited ability to capture the complexity of patient data, often relying on a narrow set of variables. More robust approaches, such as machine learning and statistical modeling, have demonstrated improved predictive performance by integrating diverse clinical and demographic factors. Existing models have incorporated variables such as treatment adherence, baseline viral load, CD4 count, regimen type, and socio-economic determinants. In this study, a predictive model was developed to identify patients at high risk of VF in Kenya. Key variables were selected based on their clinical relevance, and the model was validated to ensure reliability. Performance was evaluated using metrics AUCPR (precision-recall) and AUCROC (receiver operating characteristic), providing insights into its effectiveness in identifying high-risk patients. Three models—XGBoost Sparse, XGBoost Simple, and Random Forest (RF) Simple—were trained and evaluated using an imbalanced dataset, where the unsuppression rate was 9%. XGBoost Sparse outperformed other models, achieving the highest AUC-PR lift (4.93) and an AUC-ROC of 0.804, demonstrating its superior ability to detect virological failure. SHAP analysis revealed key predictors, including recent unsuppressed rate, previous viral load history, treatment failure, ART regimen optimization, and BMI. These findings emphasize the importance of targeted interventions, such as regimen optimization, adherence support, and socioeconomic assistance, to improve treatment outcomes. The study highlights the potential of ML-driven decision support tools in enhancing HIV care and reducing VF rates.

KEYWORDS: Predictive modelling, Virological failure, Electronic medical records, HIV treatment, Machine Learning, XGBoost, Random Forest, AUC-PR, AUC-ROC