Practical cardiac events intelligent diagnostic algorithm for wearable 12-lead ECG via self-supervised learning on large-scale dataset

Cardiovascular diseases have become a major global public health problem, and computer-aided diagnostic methods, particularly intelligent diagnosis, play an increasingly important role in electrocardiograms (ECGs). Transient and nonrhythmic arrhythmias are often easier to detect via continuous monitoring by wearable ECGs, and earlier treatment of these diseases is of great significance to patient health. Users often upload massive ECG data, and not all of them can be diagnosed by cardiologists, which is an opportunity and also our motivation to mine the information from these large amounts of ECGs. We collected 658,486 wearable 12-lead ECGs, among which 164,538 were diagnosed by a cardiologist and reviewed by senior cardiologists, and the remaining 493,948 ECGs were without diagnosis information. We used all ECGs to train a Siamese network via contrastive learning, then transferred the pretrained weights to the downstream classification network. Wearable ECGs are more prone to myoelectric artifacts, motion artifacts and electrode shedding than standard resting-state ECGs. To improve the robustness of the model, we designed four data augmentation operations for 1D digital multilead ECG signals. Our model can recognize 55 cardiac events, including normal rhythm, sinus rhythm, eight waveform changes and 45 common arrhythmias, and it achieved the average area under the receiver–operating characteristic curve (AUROC) and average F1 score on offline test of 0.979 and 0.618; the average sensitivity, average specificity and average F1 score during the three-month online test are 0.837, 0.961 and 0.602, respectively. Our model was evaluated on a large-scale online test set and confirmed by cardiologists to be practical in the real world. This approach can detect and recognize 55 cardiac events in a fine-grained manner, provide users with real-time intelligent diagnosis, and screen abnormal fragments in long-term ECG monitoring for further diagnosis by cardiologists.