MACHINE LEARNING-BASED PREDICTION OF HEART FAILURE MORTALITY USING CLINICAL BIOMARKERS AND IMBALANCED DATA STRATEGIES

Abstract

Heart failure continues to account for high global mortality rates despite improvements in cardiovascular medicine. Early mortality risk detection is essential for healthcare resource planning and personalised intervention. This study develops a machine learning framework for mortality prediction in patients with chronic heart failure, using demographic, clinical and biochemical data. A dataset containing 299 patient records and 13 biomarkers was pre-processed, normalised, and evaluated to identify feature-outcome relationships. Several modelling strategies were assessed, including Random Forest and Logistic Regression. Class imbalance was handled using Tomek Links, SMOTE and Random Oversampling to reduce decision boundary distortion. Random Forest trained on Tomek Links rebalanced data produced the strongest performance, achieving test accuracy of 0.91 and a ROC-AUC of 0.94. Logistic Regression achieved lower accuracy (0.82) and weaker minority recall. The results highlight the effect of biomarker depth on model separability, where serum creatinine, ejection fraction and age were the strongest predictors. The outcomes demonstrate that machine learning models can effectively identify high-risk patients and that minority data resampling significantly improves mortality classification reliability. The approach offers potential value for clinical decision support systems and prioritised care pathways.