Background: Manual rhythm classification in 12-lead ECGs is time-consuming and operator-biased. We present an automatic algorithm for rhythm classification using CinC’s challenge 2020 dataset (team: Germinating).
Methods: Four classifiers were implemented to recognize atrial fibrillation (AF), first-degree atrioventricular block (I-AVB), Left/Right bundle branch block (LBBB/RBBB), premature atrial/ventricular complex (PAC, PVC) and ST-segment depression/elevation (STD/STE) occurrences.
Firstly, R peaks were detected. Fixed windows containing ECG waves and segments of clinical relevance, as P-wave, QRS-complex, PQ-ST (PT-segment with QRS-segment removal), T-wave, PR, RTend were extracted. For each wave/segment, subject-specific templates were computed. Maximum cross-correlation values and standard deviation of intra-subject windows against subject-specific templates were computed. Furthermore, an inter-subject rhythm-specific template was computed for each rhythm and wave/segment using the patient-specific templates. Cross-correlation between the subject-specific templates and the inter-subject templates was computed for all waves/segments. Maximum values and standard deviations were stored as features. All classifiers considered also heart hate variability metrics, age and gender.
The first classifier is a support vector machine (SVM) that uses the features related to P-wave and PR-segment templates, representing atrial activity data, to detect AF and I-AVB. The second, is an SVM with ventricular activity-related features to detect LBBB, RBBB, STD, STE, using the information calculated on the QRS-complex, PQ-ST segment, RTend-segment and T-wave from the comparison with the rhythm-specific templates. The third and fourth classifiers are bagged trees fed with the P-wave template-related intra-patient features for PAC and PVC detection as well as RR, P-wave and QRS of the 5 beats with shorter RR-intervals information. A ten-fold cross validation was performed. Inter-patient models were built at each k-fold using only training data.
Results: The first results of less complex version than the one described obtained: F-score=0.463, G-score=0.223. Results of the most updated and non-submitted version after tenfold cross validation had mean Fbeta-score=0.663±0.015 and Gbeta-score=0.450±0.018.