Automatic Detection for Multi-Labeled Cardiac Arrhythmia Based on Frame Blocking Preprocessing and Residual Networks

Figure 1. Flow chart diagram of the algorithm for multi-type cardiac arrhythmia classification.

Figure 2. Illustration of frame blocking for pre-processing ECG signal. (A) Method of frame blocking. (B) Example of 12-lead ECG data segments after frame blocking processing.

Figure 3. Diagram of the structure of dense block1 and dense block2. BN, batch normalization; ReLU, rectified linear units; Conv1D, one-dimension convolutional layer.

Figure 4. Structure of Attention Mechanism. Correlation between the key-value pairs of the input time sequences and the query (a condition value) is evaluated, based on which the weight of each value is calculated. Through the weighted summation, the attention value for each element in the input time sequence can be assign.

Figure 5. Structure of the proposed neural network.

Figure 6. Comparison of F1 scores between different models based on the same test samples. F1 scores of the proposed model show the best performance of the model as compared with others with values of 0.959 for AF, 0.937 for an intrinsic paroxysmal atrioventricular block (I-AVB), 0.958 for LBBB, 0.885 for Normal, 0.848 for PAC, 0.920 for PVC, 0.965 for RBBB, 0.841 for STD, and 0.868 for STE. Specific values of F1 scores of the other three models are shown in Supplementary Table 7.

Figure 7. Comparison of overall F1 scores between using the proposed block framing and the common padding method for pre-processing ECG data for classification.

Figure 8. Comparison of performance between the proposed model and the Challenge-best model tested on the CPSC 2020 dataset for various types of arrhythmias. F scores of the proposed model are 0.940 for AF, 0.856 for intrinsic paroxysmal atrioventricular block (I-AVB), 0.898 for LBBB, 0.870 for Normal, 0.743 for PAC, 0.798 for PVC, 0.922 for RBBB, 0.841 for STD, and 0.868 for STE. Comparison of the F1 score between them is listed in Supplementary Table 8.

Figure 9. Performance of the proposed algorithm on PTB XL dataset for five diagnosis labels. F scores of each label are 0.853 for NORM, 0.852 for MI, 0.842 for STTC, 0.853 for CD, and 0.791 for HYP.