Classification of AF and other arrhythmias from a short segment of ECG using dynamic time warping

Atrial Fibrillation(AF) is a major public health risk but its identification is challenging because it may be episodic and non-symptomatic. Automatically identifying episodes of AF from a short segment of ECG would, thus, be beneficial. As a response to the Physionet/Computing in Cardiology Challenge 2017 we have implemented a three-stage classifier which can classify segments of ECG into Noisy, Normal, AF or Other Rhythm. We employ a state-of-the-art SQI for identifying noisy segments and then learn two different Support Vector Machine (SVM) classifiers using features extracted from the ECG. The features used are derived using a template matching approach via Dynamic Time Warping and using the statistical characteristics of the R-R intervals. Our average F1 score on the validation set was 0.66.