Deep learning for the classification of atrial fibrillation using wavelet transform-based visual images.

Background: As the incidence and prevalence of Atrial Fibrillation (AF) proliferate worldwide, the condition has become the epicenter of a plethora of ECG diagnostic research. In recent diagnostic methodologies, Morse Continuous Wavelet Transform (MsCWT) is a feature extraction technique utilized to draw out distinctive attributes of ECG signals. In our study, we explore the employment of MsCWT in the classification of AF with ECG signals in a continuum.

Results: We present a MsCWT image-based deep learning machine for AF differentiation. For the training, validation, and test sets, we achieved average accuracies of 97.94%, 97.84%, and 91.32%; and overall F1 scores of 97.13%, 96.86%, and 89.41% respectively. Moreover, AUC ROC curves of over 0.99 were obtained for all classes in the training and validation sets; and were over 0.9679 for the test set.

Conclusions: Training deep learning machines for the classification of AF with MsCWT-based images demonstrated to yield favorable outcomes and achieved superior performance amongst studies utilizing the same dataset. Though minimal, the conversion of signals into wavelet form with MsCWT may drastically improve outcomes not only in future ECG signal studies; but all signal-based diagnostics.