Cardiovascular Disease Detection: Atrial Fibrillation Analysis Using Transfer Learning

Abstract

Atrial fibrillation is a common heart rhythm problem. It increases the risk of stroke, heart failure, and other serious issues. This study used deep learning to sort AF and non-AF ECG signals. It turned one-dimensional waveforms into image-based representations. A dataset of 125,000 ECG samples, drawn from the MIT-BIH and PTB Diagnostic ECG databases was used. The processing is done. After this, it explores three major CNN architectures, SqueezeNET, AlexNET and Inceptionv3 and applies a transfer learning approach. Some of the pre trained layers of each model are unfrozen. This in turn helps extract the features which are crucial for the AF detection. In addition, dense layers were added to enhance classification. Inceptionv3 shows an accuracy of 88%. AlexNet and SqueezeNet both reach 76%. The results showed that CNN based transfer learning is a powerful tool to identify irregular heartbeat in clinical practice. So quick and accurate diagnostics is the thing here. Cardiac arrhythmia detection can be done faster with automated frameworks. Better patient outcomes result when manual ECG interpretation is reduced. This is an alteration permitting earlier intervention. Three key areas of future work will see the dataset extended, the design of network enhanced, and methods of interpretation refined. These steps will assist in improving the automated AF detection systems.