A novel deep-learning method for fibrillatory waves extraction

Atrial fibrillation (AF) is the most common sustained supraventricular arrhythmia caused by a dysfunction of the sinus atrial node, which is no longer able to guide atrial depolarization and, consequently, ventricular depolarization. As a result of this, atrial contractility is lost causing an inability to completely empty blood from atrial appendage leading to the risk of clot formation and subsequent thromboembolic events. The ECG analysis represents the most well-established noninvasive technique used to detect atrial fibrillation. Most of the works on the analysis of ECG records for AF detection are based on heart rate variability, i.e., on the R-R intervals, even if, the most relevant information in atrial arrhythmias is contained in the fibrillatory waves (F-waves) which replace the ordinary P-waves, related to the depolarization of the atria. Due to their stochastic shapes and little amplitude, the process of detection, extraction and visual inspection by clinicians of the F-waves represents a really challenging task. Classical techniques used to carry out these tasks involve the use of different signal processing principles such as principal component analysis; however, recently, deep neural networks such as 1D convolutional neural networks (CNNs), achieved high results in feature extraction and filtering of biomedical signals. For this reason, in this thesis, a two-stage deep learning method based on 1D CNNs and multipath modules is proposed to extract F-waves signals from 1-second length windows of recorded ECG of patients affected by atrial fibrillation. The system was trained and tested on a reference database, available online, for validation of methods of extraction of atrial fibrillatory waves in the ECG, which consists of records of simulated AF 12-lead ECG signals that are different combinations of real F-waves and QRST complexes. According to the results related to the testing dataset, in terms of evaluation metrics analyzed, the performances of the implemented method are really promising, with a mean correlation between the output and the target signal of 0,82 and mean values of the sum of square distances and maximum absolute distance of 0,05 au and 0,03 au, respectively. Moreover, also the mean absolute errors of the dominant frequencies and the amplitudes computed between the output and the target signals are low, with values of 0,05 Hz and 0,01 μV, respectively. To our knowledge, no other work exists in the literature which employs deep learning algorithms to extract F-waves from ECG signals of AF patients. For this reason, also according to the promising results obtained, this work can be considered a forerunner for this branch of research.