Convolutional neural networks predict the onset of paroxysmal atrial fibrillation: Theory and applications


In this study, we aimed to detect paroxysmal atrial fibrillation episodes before they occur so that patients can take precautions before putting their and others’ lives in potentially life-threatening danger.
We used the atrial fibrillation prediction database, open data from PhysioNet, and assembled our process based on convolutional neural networks. Conventional heart rate variability features are calculated from time-domain measures, frequency-domain measures using power spectral density estimations, time-frequency-domain measures using wavelet transform, and nonlinear Poincaré plot measures. In addition, we also applied an alternative heart rate normalization, which gave promising results only in a few studies, before calculating these heart rate variability features. We used these features directly and their normalized versions using min–max normalization and z-score normalization methods.
Thus, heart rate variability features extracted from six different combinations of these normalizations, in addition to no normalization cases, were applied to the convolutional neural network classifier.
We tuned the classifiers’ hyperparameters using 90% of feature sets and tested the classifiers’ performances using 10% of feature sets. The proposed approach resulted in 87.76% accuracy, 91.30% precision, 80.04% recall, and 87.50% f1-score in heart rate variability with z-score feature normalization. When the heart rate normalization was also utilized, the suggested method gave 100% accuracy, 100% precision, 100% recall, and 100% f1-score in heart rate variability with z-score feature normalization.
The proposed method with heart rate normalization and z-score normalization methods resulted in better classification performance than similar studies in the literature.
By comparing the existing studies, we conclude that our approach provides a much better tool to determine a near-future paroxysmal atrial fibrillation episode. However, although the achieved benchmarks are impressive, we note that the approach needs to be supported by other studies and on other datasets before clinical trials.