An atrial fibrillation classification method based on an outlier data filtering strategy and modified residual block of the feature pyramid network

Atrial fibrillation (AF) is a prevalent form of cardiac arrhythmia, and early detection plays a crucial role in reducing both disability and mortality rates among patients. Deep learning, as an artificial intelligence technique, offers the potential for automatic AF classification to meet the demand for timely identification. However, most existing deep learning methods demonstrate satisfactory performance on specific datasets but lack generalization capability across different datasets, posing significant challenges in the practical applications of AF classification research. To address these issues, we propose a novel outlier data filtering strategy that combines the local outlier factor and random oversampling algorithms (L-ROS). Furthermore, we introduce a modified residual block of the feature pyramid network (MRFPN). The proposed L-ROS method innovatively combines the local outlier factor algorithm with the random oversampling algorithm, effectively reducing the impact of outliers on model classification by filtering out outlier data before data enhancement. Additionally, MRFPN introduces a feature pyramid network for AF classification, which enhances model classification by integrating shallow spatial features and deep semantic features to provide richer feature information. We train our proposed method on the PhysioNet/CinC Challenge 2017 dataset and evaluate its performance on both the MIT-BIH Atrial Fibrillation (AFDB) and MIT-BIH Arrhythmia (MITDB) datasets. Compared to state-of-the-art research, our proposed method achieves significant improvements in accuracy (4.84% on the AFDB dataset; 0.36% on the MITDB dataset) and F1 metrics (4.5% on the AFDB dataset; 0.73% on the MITDB dataset), respectively.