CONTEXTUAL FEATURE NORMALIZATION ON THE PERFORMANCE OF HEART DISEASE CLASSIFICATION MODELS

Abstract

Heart disease classification models commonly employ statistical normalization techniques that standardize features according to data distribution but do not explicitly incorporate clinically meaningful cardiovascular information. This study evaluates Clinical Contextual Normalization (CCN) as an alternative feature representation strategy for heart disease classification. Using the Cleveland Heart Disease dataset (303 records), standard numerical representation and CCN were evaluated across five classifiers: Logistic Regression (LR), Support Vector Classifier (SVC), Random Forest (RF), Multilayer Perceptron (MLP), and Naïve Bayes (NB). Model performance was assessed using repeated stratified 10-fold cross-validation with five repetitions (50 evaluation folds), with recall as the primary metric because false negatives may delay clinical screening. The results revealed a classifier-dependent response to CCN. Random Forest showed a small numerical recall increase (ΔRecall = +0.0059), but the difference was not statistically significant (p = 0.6434). MLP produced the largest positive numerical recall change (ΔRecall = +0.0143) and produced 10 more aggregated true-positive predictions while false positives decreased by two, although its recall difference was also not statistically significant (p = 0.2195). In contrast, Logistic Regression showed a statistically significant recall decrease (ΔRecall = −0.0115, p = 0.0186), while Naïve Bayes exhibited the largest significant reduction (ΔRecall = −0.0333, p < 0.001). These findings demonstrate that clinically informed feature representation does not uniformly improve predictive performance but produces classifier-dependent effects. Further validation using larger and more diverse datasets is required before clinical deployment.