The body surface electrocardiogram (ECG) represents a global measure of the electrical heart activity. Acute cardiovascular events are known to be strongly modulated by a variety of physiological factors and thus of major interest for the early detection of cardiac abnormalities. Cardiac abnormalities result in a disturbed ECG and can be identified by visual inspection or automated signal processing techniques. Especially the latter open up new possibilities to support the daily work of clinicians and extend diagnostic possibilities. The objective of this study was to investigate the capabilities of morphological ECG features for the automatic classification of 12-lead ECGs.
This study was conducted as part of the PhysioNet Challenge 2020 (team name IBMTpeakyFinders). The data of the challenge was recorded by multiple sources and covers normal sinus rhythm and eight abnormal types. The entire dataset consisted of 6,877 12-lead ECG recordings with a duration of 6 to 60 seconds.
To take into account the ECG waveform morphology we focused on signal processing techniques based on template delineation that yield the template’s fiducial points. With these features beat and beat-to-beat intervals were calculated in addition to features based on the whole ECG waveform. Our classification approach utilized a decision tree ensemble and the light gradient boosted machine framework. To classify multiple types of abnormalities we used a One-Vs-Rest approach. Performance was evaluated with a 4-fold stratified cross-validation.
A first test implementation of our approach achieved a F2 score of 0.123, a G2 score of 0.055, and a geometric mean of 0.082 with a runtime of 16 minutes. This implementation is limited to a small subset of features. By integrating the entire feature set the full power of our approach should evolve to purpose the goal of improving diagnoses of cardiovascular diseases.