DEEP LEARNING AND ENSEMBLE MODELS FOR AUTOMATED ECG ARRHYTHMIA CLASSIFICATION

Abstract

Electrocardiography (ECG) serves as the primary diagnostic method which medical professionals utilize to identify heart problems and track cardiac function. The process of reading ECG signals requires extended time from doctors because it involves extensive data analysis which increases the probability of mistakes. The progress of artificial intelligence and deep learning technology has led to the creation of automated systems which can identify cardiac arrhythmias with exceptional precision. This research develops an automated ECG beat classification system which uses deep learning to process ECG signals through hybrid systems that employ convolutional neural networks (CNNs) and bidirectional long short-term memory networks (BiLSTM) and attention mechanisms and residual learning methods. The ECG dataset which the system uses includes four different heartbeat types: Normal (N) and Supraventricular Ectopic Beat (SVEB) and Ventricular Ectopic Beat (VEB) and Fusion Beat. Each heartbeat segment contains 160 normalized signal samples which represent the complete ECG waveform. The dataset creates training and validation and testing sets by using stratified sampling method which ensures equal class distribution across all groups. The research team tested three different deep learning models which they designed to compare with three different systems: ECG-ResNet and CNN-BiLSTM and CNN-BiLSTM with Attention. The research findings prove that hybrid systems achieve superior performance because they successfully capture both the spatial and temporal patterns present in ECG data. The CNN-BiLSTM-Attention model achieved the highest accuracy of approximately 97%, followed by CNN-BiLSTM with about 96%, while ECG-ResNet achieved around 92% accuracy. The assessment of performance used three main metrics which included accuracy and precision and recall.