Application of Dense Neural Networks for Detection of Atrial Fibrillation and Ranking of Augmented ECG Feature Set

Figures

• 

  Figure 1. General DenseNet architecture optimized for rhythm classification {N,AF,O,X} with 137 input ECG features, batch normalization layer and configurable dense layers (depth, size): {number dense layers} = [1–4], number nodes@N{1–3} = [0–1024] (red dots). W{1–4} stand for all trainable dense layer weight matrices.

• 

  Figure 2. Illustration of the process for computation of Case Weights Importance and Global Weights Importance based on an example neurons weights matrix Wi∈ℝ1×40 and neurons activation state matrix αiXk∈ℝ11×40.

• 

  Figure 3. Performance of all DenseNet architectures in the grid search, grouped to the number of dense layers: (a) F1 (Total) score observations in respect to the number of trainable parameters; (b) Statistical distributions of F1 (Total) score, showing the median value (line), interquartile range 25–75% (boxes), non-outlier range (whiskers) and individual observations (markers).

• 

  Figure 4. Performance of all NN architectures in the grid search with two (DenseNet-2), three (DenseNet-3) and four (DenseNet-4) dense layers in function of the number of nodes of the last hidden layer.

• 

  Figure 5. Influence of rhythm class weights W (Normal, AF and Other rhythms) on DenseNet-3@128-32-4 performance. The contour plots of F1-score measured within W grid [0; 0.5] justify the choice of the weights as the average values in the mid-area of maximal performance (+mark): W(Normal) = 0.23, W(AF) = 0.25, W(Other) = 0.3.

• 

  Figure 6. DenseNet-3 performance dependency on batch size (a) and learning rate (b).

• 

  Figure 7. Global Weights Importance presented for 137 input features and four rhythm classes. The latter are used to construct stacked barplots, which present the mean absolute values computed by 20 × 5-fold independent runs of DenseNet-3@128-32-4 model. The values are reported without normalization.

• 

  Figure 8. Performance evaluation in case of feature reduction based on Global Weights Importance rank for: (a) Different architectures of DenseNet; (b) DenseNet-3@128-32-4 performance for different rhythms.

• 

  Figure 9. Feature rank given to 137 input features in descending order of their Relative Feature Importance to detect specific rhythm class. The dots and whiskers represent the mean absolute values and standard deviations of the Relative Feature Importance computed by 20 × 5-fold independent runs of DenseNet-3@128-32-4 model.

• 

  Figure 10. Visual representation of signed Relative Feature Importance for the top-20 features ranked in Table 6 to maximally contribute to the detection of each of the four rhythm classes. The bars and whiskers represent the mean values and standard deviations of the Relative Feature Importance computed by 20 × 5-fold independent runs of DenseNet-3@128-32-4 model. Signed values of Relative Feature Importance are normalized in the range [−1; 1]. They comprehensively indicate the value of the feature that would contribute the positive detection of specific class, i.e., positive sign specifies high feature value and negative sign specifies low feature value.

• 

  Figure 11. ECG example (file A00060) of correctly classified Normal rhythm. The single-lead ECG strip is representative to sinus rhythm corrupted with baseline and high-frequency artifacts due to poor electrode contact or/and patient’s movements. No discernible repeating P-waves are present, and RR intervals seem ‘irregularly irregular’. The barplots represent the Case Feature Importance of the top-10 ranked features, which maximally contribute to the output classification probability (p) of four rhythms. The ECG strip is definitively detected as Normal rhythm (P = 0.955) based on HRV features, P-wave presence and QRS-width by DenseNet-3@128-32-4 model.

• 

  Figure 12. ECG example (file A01233) of erroneously classified Normal rhythm as AF. The single-lead ECG strip is representative to peak artifacts and significant variation of RR intervals in some parts of the record. Visual diagnosis of other dysrhythmias such as atrial flutter and multifocal atrial tachycardia cannot also be safely excluded. The barplots represent the Case Feature Importance of the top-10 ranked features, which maximally contribute to the output classification probability (P) of four rhythms. The ECG strip is definitively detected as AF (p = 0.847) based on HRV and atrial activity estimation, although Other rhythm (P = 0.138) is also weakly indicated by DenseNet-3@128-32-4 model.