Explainable detection of atrial fibrillation using deep convolutional neural network with UCMFB

Atrial fibrillation (AF) is considered to be the most dangerous cardiovascular disease and its prevalence is growing year by year. In this work, an automated detection system for the early identification of atrial fibrillation is presented. A deep convolutional neural network (DCNN) model along with the uniform cosine modulated filter banks (UCMFB) is utilized for the classification purpose. The ECG signal is first decomposed into 8 sub-signals using 8-channel UCMFB and out of which first 6 sub-signals are utilised for the further processing. These sub-signals converted into images using wavelet transform packet (WTP) by considering short segments of 5 seconds. These images are then fed to DCNN model for the classification, and tested over MIT-BIH AF and normal sinus rhythm (NSR) databases. The proposed method has achieved an overall Accuracy of 99.82%, Sensitivity of 99.86%, Precision of 99.86%, Specificity of 99.87%, F1-score of 99.82%, and ROC of 100%. It is observed that the proposed method is able to achieve the best label classification when the ECG signal is converted into images. Also, the DCNN based method decreases the false diagnosis rates in identification of AF.