Objective: Electrocardiography is the most common tool to diagnose cardiovascular diseases. Annotation, segmentation and rhythm classification of ECGs are challenging tasks, especially in the presence of atrial fibrillation and other arrhythmias. Our aim is to increase the accuracy of heart rhythm estimation by the use of extreme gradient boosting trees and the development of a deep convolutional neural network for ECG segmentation.
Approach: We trained a convolutional neural network with waveforms from PhysioNet databases to annotate QRS complexes, P waves, T waves, noise and interbeat ECG segments that characterize the essences of normal and irregular heart beats. We evaluated true positive rates, positive predictive values and mean absolute differences of our annotation based on reference annotations of the QT and MIT-BIH P-wave database. Moreover, we compared the results with standard QRS detectors and Ecgpuwave. Extreme gradient boosting trees were used to determine the heart rhythm based on hand-crafted features. More precisely, a noise estimation function was used in combination with heart rate and interval data. Furthermore we defined particular features based on ECG morphology, appearance of P waves and detection of irregular beats. We examined the feature importance and identified key features for normal sinus rhythm, atrial fibrillation, alternative rhythm and noisy recordings. The classification performance was evaluated externally using F 1 scores by applying the algorithm to the hidden test set provided by the PhysioNet/CinC Challenge 2017.
Main results: The true positive rate of the convolutional neural network in detection of manually revised R peaks in the QT database was [Formula: see text] and the positive predictive value was [Formula: see text]. The detection of P and T waves reached a true positive rate of [Formula: see text] and [Formula: see text] respectively, given a 50 ms tolerance when comparing the reference to the test annotation set. The rhythm classification performance reached an overall F 1 score of 0.82 when applying the algorithm to the hidden test set.
Significance: We achieved a shared rank #9 in the post-challenge phase of the PhysioNet/CinC Challenge 2017.