Detection of Atrial Fibrillation from RR Intervals and PQRST Morphology using a Neural Network Ensemble.

Early detection and discrimination of cardiac arrhythmia, atrial fibrillation (AF) in particular, is essential for timely intervention to improve patient outcomes. In this work, an algorithm was developed to classify ECG records as normal, AF, other arrhythmia, or too noisy to classify. This algorithm, which was an entry for the PhysioNet Computing in Cardiology Challenge 2017 (the Challenge), is described. Artifact masking and QRS detection were applied to lead-I equivalent ECG records and 17 features were extracted which captured the irregularity of the RR intervals, the PQRST morphology, and artifact/noise. An ensemble of ten neural networks (NN) was trained on the features from a training set of 5,970 records. A final classification was taken by majority vote over the 10 classifiers. The trained NN models were validated on a further 2,558 ECG records and then tested on a blind out-of-sample test set of 3,658 records. A mean $F_{1}$ score across the four classes of 0.78 for the training/validation sets and 0.80 for the testing set was achieved. A higher $F_{1}$ score for the testing set indicates that overtraining did not occur, unlike most entries to the Challenge (winner mean $F_{1}$ score of 0.89 for training/validation set, and 0.83 for testing set). Performance of the Challenge winner was not ideal and there is evidence of overtraining, indicating the difficulty of classifying AF from single-lead ECG. The features and method described here performed comparably and overtraining did not occur (high likelihood of generalization) indicating a good starting point for future work.