Automatic Multi-Label ECG Classification with Category Imbalance and Cost-Sensitive Thresholding

Figures

• 

  Figure 1. The architecture of our neural network for multi-label electrocardiogram (ECG) classification. The input of the network is an ECG signal in the shape of (b × n × l), where b is the batch size, n is the number of sampling points, and l is the number of ECG leads. The input ECG is first processed by a residual neural network to extract a feature map. The feature map is then processed by an attention layer to extract feature vectors. The inputs of the attention layer include keys (K), values (V), and queries (Q). The attention layer outputs a feature vector (denoted by FVi) for each category (indexed with i). m is the number of categories that are queried. FVi is finally processed by a dedicated fully connected (FC) layer with sigmoid activation to generate the prediction for category i (denoted by pi).

• 

  Figure 2. The cost-weighted accuracy (CWAcc) scores of the model under different values of α. Each line in this figure corresponds to a model trained in a certain cross-validation fold.

• 

  Figure 3. The comparison between scores with and without adaptation to new cost definitions in 100 simulated cost changes. The x-axis indicates the scores that are achieved by CICST using the original cost definitions, while the y-axis indicates the scores that are achieved by CICST using the updated cost definitions. Each solid circle in the figure represents a simulated cost change.

• 

  Figure 4. Comparisons of the scores achieved by different thresholding methods after the costs are changed. (a) Scores of different thresholding methods after 50% of the entries of the benefit matrix are increased by 50%; (b) Scores of different thresholding methods after 50% of the entries of the benefit matrix are decreased to 70% of the original values.

• 

  Figure 5. Instances of ECG recording and model predictions. The subgraph on the left of each row shows the ECG signal (including lead II and lead V2), while the subgraph on the right shows the model predictions for the left ECG signal. The predicted probabilities for the physician-annotated labels are in red bars and marked with red downward triangles above the bars, while the predictions for other labels are shown in the cyan bars. The thresholds calculated by CICST are shown in blue dashed lines. Four instances are shown in this figure: (a) an ECG recording with annotated labels of atrial fibrillation (AF) and right bundle branch block (RBBB); (b) an ECG recording with labels of sinus tachycardia (STach) and T-wave abnormal (TAb); (c) an ECG recording with labels of sinus bradycardia (SB) and TAb; and (d) an ECG recording with labels of AF, RBBB, and premature ventricular contraction (PVC). This figure also shows the predicted probabilities for other labels, including atrial flutter (AFL), left bundle branch block (LBBB), first-degree atrioventricular block (IAVB), normal sinus rhythm (NSR), premature atrial contraction (PAC), and T-wave inversion (TInv).