Estimation of atrial fibrillation using arbitrary normal ECG segments based on convolutional neural networks

In this study, we proposed a novel method to estimate atrial fibrillation (AF) based on convolutional neural networks (CNNs). The CNNs model consisted of five layers including convolutional and max-pooling layer. Dropout and batch-normalization were performed to prevent overfitting. For performance evaluation, arbitrary normal ECG segments were used from MIT-BIH Normal Sinus Rhythm Database (nsrdb) and Atrial Fibrillation Database (afdb). The ECG signal was preprocessed by moving average filter (2nd order) and discrete wavelet transform (DWT) was used to remove the baseline wander and spike noise. The result showed the accuracy of 99.1%, sensitivity of 99.8%, and specificity of 98.5%, respectively, for the test set.