Statistical Feature Extraction Based on Wavelet Transform for Arrhythmia Detection

Abstract

Early detection of arrhythmia through electrocardiogram (ECG) signals is crucial for preventing severe cardiac conditions. This study proposes a binary classification approach using statistical features derived from wavelet-transformed ECG signals. The MIT-BIH Arrhythmia Database was used, with signals filtered using a 0.5–50 Hz Butterworth bandpass filter. Signals were segmented into 360-sample windows with 100-sample overlap and labeled based on the majority annotation within each window. Wavelet transformation using Symlet 8 at level 4 was applied, followed by the extraction of eight statistical features: mean, standard deviation, variance, skewness, kurtosis, interquartile range (IQR), root mean square (RMS), and zero crossing rate (ZCR). These features were classified using MLP, KNN, and SVM models. MLP and KNN achieved the highest accuracy of 92.46%, while SVM had lower accuracy (72.99%) but high recall (94.21%). The results demonstrate the effectiveness of wavelet-based statistical features for lightweight and accurate arrhythmia detection.