xECGNet: Fine-tuning attention map within convolutional neural network to improve detection and explainability of concurrent cardiac arrhythmias

•Attention fine-tuning improves multilabel classification and explainability of CNN.•This is achieved by simply adding an L2-norm to the objective function.•Subset accuracy is proposed as a new criterion of multilabel ECG classification.•xECGNet achieves state-of-the-art multilabel subset accuracy of 84.6%.•xECGNet yields visual explanations that reflect features of multiple labels. Background and objectiveDetecting abnormal patterns within an electrocardiogram (ECG) is crucial for diagnosing cardiovascular diseases. We start from two unresolved problems in applying deep-learning-based ECG classification models to clinical practice: first, although multiple cardiac arrhythmia (CA) types may co-occur in real life, the majority of previous detection methods have focused on one-to-one relationships between ECG and CA type, and second, it has been difficult to explain how neural-network-based CA classifiers make decisions. We hypothesize that fine-tuning attention maps with regard to all possible combinations of ground-truth (GT) labels will improve both the detection and interpretability of co-occurring CAs. Methods To test our hypothesis, we propose an end-to-end convolutional neural network (CNN), xECGNet, that fine-tunes the attention map to resemble the averaged response maps of GT labels. Fine-tuning is achieved by adding to the objective function a regularization loss between the attention map and the reference (averaged) map. Performance is assessed by F1 score and subset accuracy. Results The main experiment demonstrates that fine-tuning alone significantly improves a model’s multilabel subset accuracy from 75.8% to 84.5% when compared with the baseline model. Also, xECGNet shows the highest F1 score of 0.812 and yields a more explainable map that encompasses multiple CA types, when compared to other baseline methods. Conclusions xECGNet has implications in that it tackles the two obstacles for the clinical application of CNN-based CA detection models with a simple solution of adding one additional term to the objective function.