Ranking of the most reliable beat morphology and heart rate variability features for the detection of atrial fibrillation in short single-lead ECG

Abstract

Objective: This study participated in the 2017 PhysioNet/CinC Challenge dedicated to the classification of atrial fibrillation (AF), normal sinus rhythm (Normal), other arrhythmia (Other) and strong noise, using single-lead electrocardiogram (ECG) recordings with a duration <60 s. The aim is to apply a linear threshold-based strategy for arrhythmia classification, ranking the most powerful time domain ECG features that could be easily reproduced on any platform. Approach: An algorithm for time domain ECG analysis was designed to extract 44 features with focus on the following: noise detection; heart rate variability (HRV) analysis; beat morphology analysis and delineation of P-, QRS-, and T-waves in the robust average beat; detection of atrial activity by the presence of P-waves in the average beat and atrial fibrillatory waves (f-waves) during TQ intervals. A linear discriminant analysis (LDA) classifier was optimized on the Challenge training set (8528 ECGs) by stepwise selection of a nonredundant feature set until maximization of the Challenge F1 score. Heart rate (HR) was an independent factor for the LDA classifier design, particular to bradycardia (HR ⩽ 50 bpm), normal rhythm (HR = 50–100 bpm), tachycardia (HR ⩾ 100 bpm). Main results: The algorithm obtained official Challenge F1 scores of 0.80 (Overall), 0.90 (Normal), 0.81 (AF), 0.70 (Other), and 0.54 (Noise) on the hidden Challenge test set (3658 ECGs). This is equivalent to a true positive rate (TPR) = 90.1% (Normal), 81.5% (AF), 67.7% (Other), and 69.5% (Noise), and a false positive rate (FPR) = 13.6% (Normal), 2.3% (AF), 7.7% (Other), and 1.5% (Noise). Significance: The top five features, which together contributed to about 94% of the maximal F1 score were ranked: (1) proportion of RR intervals differing by >50 ms from the preceding RR interval; (2) Poincaré plot geometry estimated by the ratio of the minor-to-major semi-axes of the fitted ellipse; (3) P-wave presence in the average beat; (4) mean percentage of the RR interval first differences; and (5) mean correlation of all beats against the average beat. The global rank of feature extraction methods highlighted that HRV alone was able to provide 92.5% of the maximal F1 score (0.74 versus 0.8). The added value of more complex ECG morphology analysis was less significant for Normal, AF, and Other rhythms (+0.02 to 0.08 points) than for Noise (+0.19 points); however, these were indispensable for wearable ECG recording devices with frequent artefact disturbance.