Atrial Fibrillation Detection with Spectrogram and Convolutional Neural Networks

Atrial fibrillation (AF) is one of the most common heart arrhythmias and can lead to various complications such as heart failure, stroke, reduced exercise capacity, palpitations, anxiety, shortness of breath, and high blood pressure if not diagnosed promptly. In this study, we investigated the application of time-frequency domain techniques and artificial intelligence tools for the diagnosis of AF. We proposed two custom-designed Convolutional Neural Network (CNN) architecture. 24-hour Holter ECG records from patients with AF and control subjects from the Cardiology Department of Ege University were used as dataset. Ten seconds of ECG time series signals were employed to train a 1D CNN, while spectrogram images created from these signals were used to train a 2D CNN. We observed that the proposed spectrogram-2D CNN outperformed the 1D CNN, benefiting from the time-frequency information extracted by the spectrogram.