Identification of features for machine learning analysis for automatic arrhythmogenic event classification

Cardiac arrhythmias are the leading cause of death in the western world, where atrial fibrillation (AF) is the most common arrhythmias. The PhysioNet/CinC 2017 Challenge aimed to trigger a design of an algorithm that accurately classifies short single ECG lead record to 4 categories: normal rhythm, atrial fibrillation, noisy segment or other arrhythmias. The algorithm was optimized on randomly selected records out of the challenge learning set (8528 records after reassuring it includes 60.43% of normal records, 0.54% of noisy records, 9.04% of AF records and 30% of other rhythm disturbance) and tested on hidden test database. A novel R peak detector was used to accurately detect the R peaks. Based on the R peak annotation, the P,Q,S and T peaks were detected and ECG beat morphology was extracted. Quadratic SVM classifier that include combination of 62 features was used to classify the short ECG record to one of the four categories mentioned above. For records which were classified as "normal" additional neural network classifier was applied. Our algorithm reached results of total score (F1) of 0.8 (ranked 24 out a total of 90 open-source software entries), whereas normal rhythm score (F1n) was 0.9, AF rhythm score (F_1a) of 0.81, and other rhythm score (F_1o) of 0.69.