SERN-AwGOP: Squeeze-and-Excitation Residual Network With an Attention-Weighted Generalized Operational Perceptron for Atrial Fibrillation Detection

Atrial fibrillation (AF) is a prevalent cardiac arrhythmia associated with increased stroke risk. Early and accurate detection of AF is crucial for effective management and improved patient outcomes. This paper presents SERN-AwGOP, a deep learning model designed for robust and efficient AF detection using 12-lead electrocardiogram (ECG) signals. The model employs branched nodal operators with dilated and separable convolutions, capturing multi-scale features across the 12-lead input. Channel-wise recalibration and an attention mechanism dynamically weight features within each residual block. For AF detection, SERN-AwGOP achieves a 5-fold mean accuracy of 99.11% on the CUSPH database and 93.90% on the CPSC 2018 database. An occlusion study highlights the P-wave and QRS complex as critical ECG segments for AF identification. The model demonstrates robustness against common ECG signal distortions, maintaining 93.03% accuracy on data affected by Gaussian noise, time warping, and random baseline wander. It achieves 97.78% accuracy on hardware-corrupted data with electrode disconnections and amplifier saturation. SERN-AwGOP exhibits high data efficiency, achieving 99.31% accuracy with only 25% of the training data, and shows potential for cross-database generalization with 73.14% accuracy when trained on CPSC 2018 and tested on CUSPH and even more improvements are seen through finetuning based transfer learning. These results indicate SERN-AwGOP’s resilience to real-world ECG signal processing challenges and its potential for practical clinical applications in AF detection.