Detection and Semantic Segmentation of Atrial Fibrillation Signals Using U-Net Model

Abstract

An abnormal heart rhythm called atrial fibrillation is a dysfunction in the cardiacconduction system that is often life-threatening or reduces the quality of life. This study aims to develop a custom-designed artificial intelligence model UNet to determine whether patients have atrial fibrillation (AF) disease. The purpose is fast and accurate detection. The MIT-BIH Atrial Fibrillation dataset available on the Kaggle platform was used for performance evaluation. The MITBIH Atrial Fibrillation Database includes 25 long-term ECG recordings, however in this project only 18 patient recordings are used due to some of the recordings being unreadable. Two channel ECG signals recorded at 250 samples per second are included in the individual recordings from 18 patients each lasting 10 hours. All simulations were implemented with Phyton in the Kaggle environment. To get the best result the model structure of our U-Net model and parameters such as the activation function and batch size were selected heuristically. Several experiments were done and the model's performance was observed for the different loss functions optimizers and metrics. Our system reaches an accuracy of 93.91% a precision of 99.86% and a recall of 82.40% which results in an F1 score of 90.30%. © 2023 Elsevier B.V. All rights reserved.