Generalizable Beat-by-Beat Arrhythmia Detection by Using Weakly Supervised Deep Learning

Figures

• 

  FIGURE 1. Comparison between different methodologies of machine learning for arrhythmia detection. (A) The supervised learning (SL) method for heartbeat/segment classification which uses fine-grained annotations for model training and achieves fine-grained predictions. (B) The SL method for ECG recording classification which uses coarse-grained labels as supervision and achieves coarse-grained predictions. (C) The weakly supervised learning (WSL) method for beat-by-beat classification which uses coarse recording labels for model training but achieves fine-grained predictions. N = normal or bundle branch block beat. SVEB = supraventricular ectopic beat. VEB = ventricular ectopic beat.

• 

  FIGURE 2. The schematic diagram of the weakly supervised deep learning framework for arrhythmia detection (WSDL-AD). CNN = convolutional neural network. N = normal or bundle branch block beat. S = supraventricular ectopic beat. V = ventricular ectopic beat.

• 

  FIGURE 3. The results of one-way ANOVA with multiple pairwise comparison for the knowledge-based features. (A) The results for relative RR interval. (B) The results for RR entropy. The symbol “****” indicates that the corresponding two groups are significantly (p < 0.0001) different from each other in features values. N = normal or bundle branch block beat. S = supraventricular ectopic beat. V = ventricular ectopic beat.

• 

  FIGURE 4. The structure of residual convolutional network (ResNet) for feature learning. Res block = residual block. Conv = convolutional layer. Batch Norm = batch normalization. ReLU = rectified linear unit.

• 

  FIGURE 5. Precision-recall curves (PRCs) of the detection for ventricular ectopic beats (VEB) and supraventricular ectopic beats (SVEB) on the total dataset in different learning settings. (A) PRCs of VEB detection. (B) PRCs of SVEB detection. SL = supervised learning. WSL = weakly supervised learning. AP = average precision.

• 

  FIGURE 6. Distributions of the F 1 scores on the total test data in multiple independent training sessions. (A) The histogram of the results for ventricular ectopic beat (VEB) detection. (B) The histogram of the results for supraventricular ectopic beat (SVEB) detection. SL = supervised learning. WSL = weakly supervised learning.

• 

  FIGURE 7. Precision-recall curves (PRCs) of the ablation studies on the total test data. (A) and (B) show PRCs of models with different knowledge-based features. (C) and (D) show PRCs of models with different feature extraction networks. (E) and (F) show PRCs of models with different aggregation mechanisms. GAP = global average pooling. GMP = global max pooling. LSE = Log-Sum-Exp. MGMP = masked global max pooling. AP = average precision. The parameter b of LSE is set to 5.

• 

  FIGURE 8. Qualitative results of arrhythmia detection on the MITBIH-AR-DS2 in the SL + WSL setting. (A) An example during normal rhythm. (B) An example during atrial fibrillation. (C) An example during atrial bigeminy. (D) An example during ventricular bigeminy. The vertical dashed lines indicate the positions of R peaks. In the ECG waveform charts, the reference category of each heartbeat is labeled above the ECG. In the local classification charts, the predictions for supraventricular ectopic beats (S) are drawn in blue, the predictions for ventricular ectopic beats (V) are drawn in orange, and the detected categories are labeled above the prediction lines. N denotes a normal or bundle branch block beat. Since the probabilities of N, S, and V add up to one at each sampling point, the predictions for N are not plotted in the figure for clarity.

• 

  FIGURE 9. Qualitative results of error predictions. (A) An example of misclassifying supraventricular ectopic beats (S) as normal or bundle branch block beats (N) during supraventricular tachyarrhythmia. (B) An example of misclassifying a ventricular ectopic beat (V) as an S during left bundle branch block (LBBB).