Correlation-Based Feature Selection for Enhanced Arrhythmia Classification

The accurate diagnosis of cardiac arrhythmias relies on analyzing patterns in electrocardiogram (ECG) signals. However, ECG datasets typically contain a large number of input features relative to the sample size, which can negatively impact model performance. This study evaluates feature selection techniques to identify an optimal subset of input features that maximizes arrhythmia classification accuracy. Eight feature selection methods encompassing filters, wrappers, and embedded approaches were compared on the UCI Arrhythmia dataset containing 279 features. The selected feature subsets were evaluated by training three classifiers - random forest, J48, and reduced error pruning tree. Results indicate that correlation-based feature selection (CFS) identified the best feature subset of 32 features, improving classification accuracy over all features by 10 Random forest achieved the highest accuracy of 86.4% with the CFS subset. Redundant and irrelevant features decreased performance for other selection methods. The analysis provides a guideline for applying feature selection in clinical applications and demonstrates methods to improve generalization and prevent overfitting in high-dimensional biomedical datasets.