Morphological Autoencoders for Beat-by-Beat Atrial Fibrillation Detection Using Single-Lead ECG

Figures

• 

  Figure 1. Morphological and rhythm changes between ECG recordings in Normal Sinus Rhythm (top) and in Atrial Fibrillation (bottom), where the absence of P-waves and the irregular heart rate are noticeable.

• 

  Figure 2. Number of AFib detection algorithm publications from 2013 to 2020, showing how the number of ANN-based algorithms have gained popularity over time. The year 2018 was boosted by the CinC2017 challenge. Data obtained from the Appendix A, Supplementary Data section of [19].

• 

  Figure 3. (a) A traditional autoencoder which is trained to reproduce the input samples, and (b) a supervised autoencoder that receives additional feedback from labeled samples.

• 

  Figure 4. The proposed model for classifying ECG heartbeat waveforms into Normal Sinus Rhythm (NSR) and Atrial Fibrillation (AFib). It consists of an encoder responsible for extracting morphological features from the ECG segments, which are then used by a classifier. The LCSD metric is also used to train the classifier to provide local rhythm information and improve classification performance.

• 

  Figure 5. Example of LCSD values for ECG beats in NSR (left) and in AFib (right). Since RR-Intervals are more regular in NSR, their consecutive differences are smaller compared to AFib.

• 

  Figure 6. Preprocessing steps of the ECG recordings.

• 

  Figure 7. (a) Example of second R-peak removal from an AFib ECG segment using zero-padding. (b) Use of DMEAN to detect outliers in a set of ECG segments from a recording—valid segments are depicted as solid lines, while outliers are displayed in dashed lines. Data from the CinC2017 database [23].

• 

  Figure 8. Distribution of LCSD values for NSR and AFib in two databases. The Mann–Whitney U test revealed significant statistical differences between the distributions, with p-values smaller than 0.0001 (marked by ****).

• 

  Figure 9. Comparison of Encoder-MLP (left) and Supervised Autoencoder (center) on CinC2017 data shows that SupAEs have insignificant changes in convergence and performance in terms of lower classification losses. The ROC curve of SupAE with and without the LCSD metric (right) highlights the best threshold value.

• 

  Figure 10. Comparison of Encoder-MLP (left) and Supervised Autoencoder (center) on AFDB data shows the SupAE’s quicker convergence and better performance in terms of lower classification losses. The ROC curve of SupAE with and without the LCSD metric (right) highlights the best threshold value.

• 

  Figure 11. Comparing the performance of Encoder-MLP and Supervised Autoencoder on AFDB data using a patient-based split. The learning curves on the (left) and (center) show that both models face difficulties in converging the validation loss, possibly due to the presence of unseen pathological waveforms. The ROC curve of the SupAE model (right) with and without the LCSD metric highlights the best threshold value.

• 

  Figure 12. Portion of a pathological ECG waveform from the AFDB database (recording 04043).