Regularized regression models for accurate prediction of HIV progression in ART patients: a comparative study

Abstract

This study compared four regularized regression techniques Ridge, Least Absolute Shrinkage and Selection Operator (LASSO), Adaptive LASSO, and Elastic Net to analyze survival models for HIV progression. The HIV patients were those under antiretroviral treatment, considered alongside some demographic and health variables. The selection criteria for the model with the best performance in this study were the Concordance Index (C-index) for ranking ability, Brier Score for predictive accuracy, Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), Akaike Information Criterion corrected (AICc), and Mean Square Error (MSE). The result showed that for the Ridge model, age, viral load, height, and WHO clinical stage of a patient were associated with lower survival probability (HR > 1) while sex, educational level, marital status, job status, and Body Mass Index of a patent had higher survival probability (HR < 1). LASSO and Elastic Net models showed that the height and WHO clinical stage of a patient were associated with reduced survival while the Adaptive LASSO model indicated that the WHO clinical stage is associated with lower survival probability. The analysis revealed that Adaptive LASSO had the lowest Brier Score, AIC, BIC, AICc, MSE, and highest C-index among the models evaluated when a highly collinear variable was removed. Elastic net had better performance when all predictors were included in the model, indicating that Elastic Net performed better than other models in terms of prediction accuracy and robustness against multicollinearity.