A Comparative Study of Time-Domain, Wavelet, And Morphological Features for ECG Beat Classification Using Random Forest

Abstract

Automated classification of electrocardiogram (ECG) beats can facilitate analysis of long-term cardiac recordings. This study performs a controlled comparison of statistical time-domain, discrete wavelet transform (DWT), and morphological features for three-class ECG beat classification. The MIT-BIH Arrhythmia Database was processed using a record-level training/testing protocol. Forty-eight records were considered, with 37 records assigned to training and 11 reserved for independent testing. A maximum of 2,000 beats per class was selected from the training records, resulting in 5,999 training beats from normal (N), supraventricular (S), and ventricular (V) classes. The independent test set retained its original distribution and contained 19,207 beats. Ten statistical time-domain features, 24 DWT features, and eight morphological descriptors were extracted from 270-sample ECG beat segments. Random Forest, support vector machine (SVM), and XG Boost classifiers were evaluated. Five-fold grouped cross-validation was performed using record identity as the grouping variable.The wavelet-morphology representation with Random Forest achieved 83.84% accuracy, 69.02% balanced accuracy, and 58.61% macro F1-score on the held-out test set. The class-wise F1-scores for N, S, and V were 90.48%, 10.66%, and 74.69%, respectively. Grouped cross-validation yielded 61.03 ± 5.34% balanced accuracy and 57.57 ± 7.42% macro F1-score. The results indicate that wavelet features provide useful multiscale information beyond the evaluated statistical time-domain descriptors, while simple morphological descriptors provide a modest additional improvement. Supraventricular beat recognition remains a significant limitation.