Cardiac Arrhythmia Detection Using Deep Learning Methods

Author

Xander Apicella, California Polytechnic State University, San Luis ObispoFollow

Available at: https://digitalcommons.calpoly.edu/theses/2969

Date of Award

3-2025

Degree Name

MS in Electrical Engineering

Department/Program

Electrical Engineering

College

College of Engineering

Advisor

Xiao-Hua (Helen) Yu

Advisor Department

Electrical Engineering

Advisor College

College of Engineering

Abstract

Cardiac arrhythmias and heart conditions are a significant cause of mortality globally. This thesis outlines machine learning (ML) methods capable of taking and classifying data taken from patients via electrocardiogram (ECG) signals. It explores two datasets, the often-studied MIT-BIH dataset and the newer and more robust CSN-ECG dataset. Three main ML model architectures were used—a simple convolutional neural network (CNN) architecture, a Residual Neural Network (ResNet) architecture, and a transformer architecture using the cutting-edge attention mechanisms its known for to classify.

Results demonstrated that the simple CNN network performed best on the MIT-BIH dataset with a 99.0% accuracy, with the ResNet-18 performing second best with a 98.5% accuracy. Only the ResNet-18 and transformer models were tested on the CSN-ECG dataset, and the ResNet-18 performed best on that dataset with a macro-averaged F1 score of 90.0%. The transformer performed poorly for both datasets due to its model complexity and overfitting. Future potential extensions of this research could be tuning the transformer, the addition of more robust data augmentation for the MIT-BIH dataset and the use of transfer learning, and the use of hybrid models for CSN-ECG database.