Atrial fibrillation detection using feedforward neural networks and automatically extracted signal features

Atrial Fibrillation (AF) is becoming an increasingly significant clinical matter as its prevalence keeps growing. Therefore, developing algorithms to accurately detect AF episodes from single-lead ECG recordings would benefit early and automatic diagnosis from monitoring devices. The aim of this study is to provide a Feedforward Neural Network (FFNN) classification model and asses its performance in order to discriminate short single-lead ECG registers among 4 groups: Normal (N), AF (A), Other rhythms (O) and noisy (~). We extracted automatically 72 features derived from ventricular activity from each of the 8528 ECG records provided by the 2017 PhysioNet/Computing in Cardiology Challenge. Next, we performed a Feature Selection (FS) and a training/validation process through a grid search over a set of FFNN training parameters. We used a F1 scoring in order to assess the classification performance. Results shown that filtering 50 features during the FS stage improved the initial classification performance from F1=0.70 to F1=0.73. The following tuning of FFNN training parameters showed the best results during the 10Fold Cross-Validation with F1=0.76 (F1n=0.87, F1a=0.78, F1o=0.65, F1p=0.45) using 200 epochs, α=0.7, β=0.0, and one hidden layer made of 128 units. The final score on the test data was F1=0.77, demonstrating the robustness of the presented method. Our strategy revealed promising classification scores using a robust validation approach. The resulting classification model is computationally low consuming during classification, hence is a good candidate to be implemented in wearable patient management systems.