Predicting Mortality Risk in Heart Failure Patients: Insights from COX Regression and Machine Learning

This study is based on baseline data and 2-year follow-up information from 145 heart failure patients in Guangxi, China, combined with a publicly available scientific dataset on Chinese heart failure patients. Multiple datasets of varying scales were constructed, and traditional Cox proportional hazards models, along with Logistic Regression, Random Forest, Support Vector Machine, XGBoost, Random Survival Forest, and Survival Support Vector Machine algorithms were employed to develop heart failure mortality risk prediction models. These models were used to identify and quantitatively evaluate both risk factors and protective factors for HF mortality. In terms of the outcome, XGBoost demonstrated superior performance on high-dimensional datasets with missing values, whereas Support Vector Machine exhibited stronger predictive capability on similarly scaled datasets without missing values. The results based on XGBoost model and evaluation based on SHAP values further confirmed that Glomerular Filtration Rate, Height and Glucose are critical predictors. For survival time analysis, the Cox models generally outperformed Random Survival Forest and Survival Support Vector Machine algorithms. The mutual validation of different modeling approaches can enhance the robustness and effectiveness of heart failure mortality risk prediction, better supporting clinical prevention and treatment decision-making.