Arrhythmia Disease Detection using SVM with Recursive Feature Elimination

Abstract

Arrhythmia is a critical cardiovascular disorder affecting approximately 1.5% to 5% of the global population. The issue of early detection remains challenging due to asymptomatic presentation and complex electrocardiogram (ECG) signal interpretation. Traditional diagnostic methods and existing machine learning approaches often struggle with high-dimensional medical data containing irrelevant features, leading to suboptimal classification performance. This study proposes an integrated approach combining Support Vector Machine (SVM) with Recursive Feature Elimination (RFE) for automated arrhythmia detection from the UCI Machine Learning Repository dataset containing 452 patient records with 278 features. The methodology incorporates comprehensive preprocessing, including normalization, Synthetic Minority Oversampling Technique (SMOTE) for class balancing, and RFE-based feature selection. Both Linear and Radial Basis Function (RBF) kernels were evaluated across four train-test split scenarios (90:10, 80:20, 70:30, 60:40). The proposed method achieved superior performance with 91.30% accuracy, 88.00% precision, 95.65% recall, and 91.67% F1-score using the RBF kernel in the 90:10 scenario. RFE successfully reduced dimensionality by 96.4%, selecting 10 optimal features from 278 original parameters while maintaining high classification accuracy. These findings demonstrate that the integration of SVM with RFE significantly enhances arrhythmia detection capability.