Myocardial Infarction Detection Based on ECG: A Combined Approach using Variational Autoencoders (VAE) and Transfer learning

Files

Singh_Bhupinder_MSc_2024.pdf.pdf (2.67 MB)

Authors

Singh, Bhupinder

Issue Date

2024

Degree

MSc Data Analytics

URI

https://hdl.handle.net/10788/4492

Publisher

Dublin Business School

Rights holder

http://esource.dbs.ie/copyright

Rights

Abstract

This work introduces the use of convolutional variational autoencoders (CVAEs) to effectively decrease the dimensionality of electrocardiogram (ECG) images without sacrificing diagnostic information, addressing the critical requirement for early and accurate myocardial infarction (MI) identification. The study uses CNN, InceptionV3, VGG19, ResNet152, and ResNet50 deep learning models to examine how resolution affects model performance. Higher resolution images improve accuracy and specificity more uniformly across models, according to the findings. VGG19 performs particularly well, despite requiring longer training cycles. Notably, the study shows that CNN is efficient and presents itself as a viable model for rapid and reliable MI diagnosis. It also reveals that CVAEs offer a balanced approach, lowering dimensionality while retaining diagnostic accuracy.

Collections

Information & Communications Technology