Abstract
In MIT-BIH Arrhythmia Dataset and PTB Diagnostic ECG. Database classification of electrocardiogram (ECG) signals is essential in timely diagnosis. Electrocardiogram (ECG) analysis is one of the most widely used methods to detect cardiac abnormalities, but manual interpretation is time-consuming and prone to variability. This creates an urgent need for automated ECG classification systems that can deliver reliable results in clinical and real-time settings. In this study, three deep learning-based models, Autoencoder with Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) are compared and evaluated, trained on mixed and matched ECG data sets using MIT-BIH and PTB Diagnostic ECG datasets that have both normal and abnormal heartbeats. The evaluation base of each model is based on an evaluation of predictive accuracy, precision, recall, F1-score, training time, inference time and model size. As results show, CNN performs best in terms of classification accuracy (66.12 %), LSTM provides quite same performance but faster in inference time (65.89 %). Autoencoder with SVM model is smaller and takes less time to train; however, its overall accuracy is 59.23 %. Such results indicate the trade-off between computation speed, accuracy of diagnosis, and thus enable to specify model choice in real-time or ECG diagnostic system with limited resources.