Classification of Cardiac Arrhythmias Using Fourier Series and Support Vector Machines

Daniel Gazzoli Nunes
Instituto Federal do Espírito Santo - Campus Vitória


Aims: This study aimed to examine the accuracy of an approach based on the fourier series for detection and classification of nine types cardiac arrhythmias: atrial fibrillation (AF), first-degree atrioventricular block, left bundle branch block (LBBB), normal sinus rhythm, premature atrial complex, premature ventricular complex, right bundle branch block (RBBB), ST-segment depression and ST-segment elevation. Methods: The training dataset consists in 6877 labeled 12-lead ECG recordings. The approach uses a segmentation algorithm to detect R-peaks in each lead, then it calculates the coefficients of the first 20 harmonics of each beat with the fourier series. Some statistics (mean, variance, etc.) are calculated over these coefficients, and also over the heart rate. These statistics are used as features in a linear support vector classifier (SVC). We trained one linear SVC for each arrhythmia, and validated them using a 5-fold cross validation. We submitted a model trained using all of the training dataset for scoring on the test data.
Results: Two scores were used to evaluate the model, the F2 score and the G2 score. The algorithm received a mean F2 score of 0.567 and a mean G2 score of 0.305 in the 5-fold cross validation, and a F2 score of 0.586 and G2 score of 0.323 on the test data. The model scored better when trying to detect RBBB, AF, and LBBB with a mean F2-score of 0.777, 0.739 and 0.618 and a mean G2-score of 0.508, 0.479 and 0.335 respectively in the 5-fold cross validation. Conclusions: The results show that this approach could produce a viable model for some of the arrhythmias we aimed to classify. However considerable improvements are needed since the accuracy is currently not satisfactory.