Loss Optimization Based Algorithm for Multi-classification of ECG Signal

There is a wide range of cardiac diseases, and the electrocardiogram (ECG) signal allows for relevant diagnostic classification. With the development of telemedicine and computer-aided diagnostic techniques, early detection and timely treatment place new demands on diagnostic algorithms. In this paper, we propose a deep learning algorithm for ECG signal multi-classification based on loss optimization, the decomposing multi-classification task of ECG signals into multiple ECG signals binary classification tasks, using the framework of a multi-task deep learning algorithm by sharing task features, and performing loss optimization in terms of both magnitude and direction of task gradients to avoid manual setting of task loss weights and negative migration due to task losses canceling each other out, thereby improving the performance of the ECG signal multi-classification task. We evaluated the proposed algorithm using the PTB-XL dataset by decomposing the ECG signals 23 classification task into 23 binary classification tasks. The experimental results showed that the macro-averaging area under the curve achieved 0.950, the accuracy achieved 0.965, the label-based macro-averaging F1 score achieved 0.583 and the sample-based F1 score achieved 0.777. The proposed multi-task deep learning algorithm showed good performance in the multi-classification of ECG signal compared to the single-task learning algorithm.