CLASSIFICATION OF ARRHYTHMIA BY USING DEEP LEARNING WITH 2-D ECG SPECTRAL IMAGE REPRESENTATION

Abstract

Automated electrocardiogram (ECG) analysis can support the screening of cardiac rhythm abnormalities when standardized preprocessing and classification are applied. This work presents an image-based ECG arrhythmia classification framework for six rhythm categories: Left Bundle Branch Block (LBBB), Normal, Premature Atrial Contraction (PAC), Premature Ventricular Contractions (PVC), Right Bundle Branch Block (RBBB), and Ventricular Fibrillation. The dataset contains 22,166 ECG waveform images divided into training and test subsets. Images are resized to 64 × 64 pixels and intensity-scaled, while shear, zoom, and horizontal-flip augmentation are applied to the training subset. A sequential CNN comprising three convolutionpooling stages, a 128-unit dense layer, dropout, and a six-unit softmax output performs multiclass classification. The trained network is stored in HDF5 and native Keras formats and deployed through a Flask-based image-upload interface. Experimental evaluation includes epoch-wise learning behavior and three interface-level classification cases. A Normal ECG is classified as Normal and mapped to a healthy-rhythm status, a Ventricular Fibrillation input is classified as Ventricular Fibrillation and mapped to an abnormal-rhythm status, and an LBBB input is classified as PVC and mapped to an abnormal-rhythm status. The results demonstrate the complete inference pipeline from ECG image input to multiclass prediction and user-facing rhythm interpretation, while also revealing class-level confusion among abnormal rhythms.