AF Identification From Time–Frequency Analysis of ECG Signal Using Deep Neural Networks

Atrial fibrillation (AF) is a life threatening cardiac abnormality having high prevalence and risk with increased rate of stroke and systemic embolism, if oral anticoagulation is not recommended. Later, this leads to morbidity and mortality. Detection of AF is challenging from the electrocardiogram (ECG) recordings, due to its complex characteristics. Manual observation of ECG is tedious, time consuming, and error prone. This manuscript proposed a novel approach for identifying AF in the presence of noise and other beats by using deep neural networks (DNN) on the 2-D patterns obtained by various time–frequency analysis methods, such as short time Fourier transform (STFT), Chirplet-transform, Stockwell-transform, and Poincare plot from 1-D preprocessed ECG recordings. The above discussed methods identify the variations due to AF in ECG. Initially, the patterns are used by the pretrained DNN models for classification. ResNet18 attained the highest accuracy of 90.56 $\%$ for the patterns of ECG obtained by Chirplet-transform on PAF prediction challenge database, whereas Chirplet patterns used by ResNet50 achieved an accuracy of 89.72 $\%$ . Based on accuracy and the number of parameters in DNN, an ensembled network is designed for improving AF classification. Ensembling of ShuffleNet and AlexNet applied over the patterns obtained by Stockwell transform achieved the highest accuracy of 93.70 $\%$ . This approach is further experimented on PhysioNet CinC 2017 challenge database, consisting of four classes (such as AF, normal, other rhythms, and noise).