TransMixer-AF: Advanced Real-Time Detection of Atrial Fibrillation Utilizing Single-Lead Electrocardiogram Signals

Atrial Fibrillation (AF) is a prevalent cardiac arrhythmia that carries an increased risk of stroke and heart failure. The manual interpretation of Electrocardiograms (ECGs) for diagnosing AF is time-consuming and subject to observer variability. Deep Learning (DL) techniques offer a promising solution for robust AF detection. This paper introduces a novel DL model that leverages ConvMixer and Transformer architecture, combining convolutional neural networks and attention mechanism to extract both local and global features from ECG data. It excels in accurately classifying ECG rhythms into two categories: normal and atrial fibrillation. Additionally, the model utilizes GradCAM++ visualization to provide interpretability. The model was evaluated using the PhysioNet/CinC 2017 Database, achieving accuracies of 91.66% on original data and 96.59% on preprocessed data. When applied to the the MIT-BIH Arrhythmia Database, it achieved accuracies of 95.86% on original data and 98.78% on preprocessed data, employing a ten-fold cross-validation strategy. This innovative approach has the potential to assist clinicians in the real-time detection of common atrial fibrillation during routine ECG screening, significantly enhancing the efficiency and accuracy of diagnosis. With a strong focus on improving patient outcomes and reducing the burden on healthcare professionals, this research represents a crucial step forward in cardiac arrhythmia detection.