On the use of a residual convolutional neural networks and images derived from Poincaré-plots for the detection of atrial fibrillation

Atrial fibrillation is the cardiac arrhythmia with the highest prevalence. Its automatic detection is an important task. This research proposes a method that separates segments presenting normal sinus rhythm from atrial fibrillation. Initially, Poincaré plots were generated for 10 consecutive RR intervals. The resulting nine points were connected by straight lines to the plot origin. These images were then presented as inputs to a residual convolutional neural network. The classification process resulted in accuracy rates up to 99.14%. This procedure was then repeated for 60 RR intervals. An accuracy rate of 99.81% was attained.