Comparative evaluation of feature selection methods for HRV-based survival modeling in HIV-positive ICU patients: a retrospective study

Heart rate variability (HRV) reflects autonomic regulation and has emerged as a promising noninvasive marker for risk stratification in critical illness. In HIV-positive intensive care unit (ICU) patients, autonomic dysfunction may influence survival, yet its prognostic potential remains underexplored. We analyzed HRV and physiological data from 145 HIV-positive ICU patients to develop machine-learning models for in-hospital survival prediction. Three feature selection techniques—correlation analysis, mutual information, and random forest importance—were systematically compared using the top 5, 10, and 15 ranked variables. Artificial neural networks (ANNs) were trained on each subset, and the most discriminative features were further evaluated through logistic regression for interpretable probability estimation. A graphical user interface (GUI) was implemented to facilitate clinical use. The correlation-based top-15 model achieved the best ANN performance (AUC = 0.90), identifying SOFA score, platelet count, and maximum heart rate as consistent predictors of survival. Random forest and mutual information approaches yielded complementary but lower discriminative power. The developed GUI integrates HRV extraction and individualized mortality prediction through a dual-tab interface. Correlation-driven feature selection produced the most accurate and parsimonious HRV-based survival models, supporting its clinical utility for real-time prognostication in HIV-positive ICU patients. The integrated ANN–logistic regression framework and GUI enhance interpretability and potential bedside deployment. Retrospective analysis; no prospective enrollment or interventions.