Predictive Modelling of cardiovascular Disease Survival Using Mutual Information and Machine Learning Across Varying Sample Sizes

In clinical data analytics, predicting survival outcomes for cardiovascular disease (CVD) is a challenging task with practical ‎implications. Using three different datasets, this study investigates how sample size affects machine learning performance ‎and generalizability. The methodology combines statistical sample-size analysis with mutual information gain, a filter-based, scalable, and domain-agnostic feature selection strategy, to identify clinically essential features. Mutual information ‎gain measures the dependency between each predictor and the target variable, ensuring computational efficiency and ‎applicability to large-scale data. Machine learning classifiers, support vector machines (SVMs) and logistic regression (LR), ‎are employed to assess predictive performance across varying population sizes. Experimental results demonstrate that ‎increasing the sample size improves model accuracy by up to 10%, recall by 5–8%, and maintains consistent specificity. ‎Furthermore, to enhance clinical reliability, the models are evaluated using the Area Under the Receiver Operating ‎Characteristic Curve (AUC-ROC), where SVM achieved an AUC of 0.965 and LR achieved 0.937, validating strong ‎discriminatory power; Also, SHAP-based feature attribution is used to improve interpretability, identifying that larger ‎sample sizes provide more stable and clinically meaningful explanations‎.