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

Figures

• 

  Fig. 1. Information for the large-scale ECG dataset. a Wearable ECG device and its standard attachment method. Artifacts are formed in ECG due to electrode displacement and poor coupling to the skin in daily life. b Frequency distribution histogram of unannotated ECGs and annotated ECGs on relative average power. For ECGs of various power, unannotated data is always more than annotated data, and fully mining the knowledge of large-scale ECGs is expected to improve the classification performance of our model. c A typical long-tail distribution of 60 diagnostic terms for ECG, and the specific number of each class is shown in Table 1. d Example I lead of good-quality ECG signal with clear and smooth physiological waveform. e Four ECG augmentation operations based on above good-quality ECG. f Two most common artifacts in wearable ECG signals: myoelectric artifact (I lead) and motion artifact (V2 lead).

• 

  Fig. 2. The performance analysis, threshold selection and extension of application of the model. a Ablation study of three strategies: pre-training weights (PW), multi-scale (MS) convolution and data augmentation operations (Aug). b Improvement of classification performance initialized by pre-trained weights compared to random weights on different number of training sets. c Improvement of classification performance by ECG augmentation operations. Each augmentation operation has a different degree of improvement, and the combination of them has the most significant improvement in classification performance. d Effect of classification performance from ECG augmentation on offline test set, good-quality set and inferior-quality set. The offline test set is divided into good-quality set and inferior-quality set according to the data quality, and the improvement of classification performance from data augmentation was most pronounced on the inferior-quality set, indicating that our model is robust to the artifacts in ECGs. e Classification performance when only partial leads are used. f Precision-recall curves of IAVB, AF and STE. g The operating point selection of class AF on the PR curve.…

• 

  Fig. 3. Neural network architectures and visualization of model’s attention. a The improved MoCo architecture for pretraining. b The architecture of our multiscale convolutional network. c Attention visualization of our model to input ECG on three arrhythmias using CAM, the red background indicates where the model pays more attention. The purple circle indicates the human expert’s definition of this arrhythmia.