Class activation map-based slicing-concatenation and contrastive learning: A novel strategy for record-level atrial fibrillation detection

Background

Deep learning-based models for atrial fibrillation (AF) detection require extensive training data, which often necessitates labor-intensive professional annotation. While data augmentation techniques have been employed to mitigate the scarcity of annotated electrocardiogram (ECG) data, specific augmentation methods tailored for recording-level ECG annotations are lacking. This gap hampers the development of robust deep learning models for AF detection.

Methods

We propose a novel strategy, a combination of Class Activation Map-based Slicing-Concatenation (CAM-SC) data augmentation and contrastive learning, to address the current challenges. Initially, a baseline model incorporating a global average pooling layer is trained for classification and to generate class activation maps (CAMs), which highlight indicative ECG segments. After that, in each recording, indicative and non-indicative segments are sliced. These segments are subsequently concatenated randomly based on starting and ending Q points of QRS complexes, with indicative segments preserved to maintain label correctness. Finally, the augmented dataset undergoes contrastive learning to learn general representations, thereby enhancing AF detection performance.

Results

Using ResNet-101 as the baseline model, training with the augmented data yielded the highest F1-score of 0.861 on the Computing in Cardiology (CinC) Challenge 2017 dataset, a typical AF dataset with recording-level annotations. The metrics outperform most previous studies.

Conclusions

This study introduces an innovative data augmentation method specifically designed for recording-level ECG annotations, significantly enhancing AF detection using deep learning models. This approach has substantial implications for future AF detection research.