Multi-stage SVM approach for cardiac arrhythmias detection in short single-lead ECG recorded by a wearable device

Use of wearable ECG devices for arrhythmia screening is limited due to poor signal quality, small number of leads and short records, leading to incorrect recognition of pathological events. This paper introduces a novel approach to classification (normal/'N', atrial fibrillation/'A', other/'O', and noisy/'P') of short single-lead ECGs recorded by wearable devices.

Various rhythm and morphology features are derived from the separate beats ('local' features) as well as the entire ECGs ('global' features) to represent short-term events and general trends respectively. Various types of atrial and ventricular activity, heart beats and, finally, ECG records are then recognised by a multi-level approach combining a support vector machine (SVM), decision tree and threshold-based rules.

The proposed features are suitable for the recognition of 'A'. The method is robust due to the noise estimation involved. A combination of radial and linear SVMs ensures both high predictive performance and effective generalisation. Cost-sensitive learning, genetic algorithm feature selection and thresholding improve overall performance. The generalisation ability and reliability of this approach are high, as verified by cross-validation on a training set and by blind testing, with only a slight decrease of overall F1-measure, from 0.84 on training to 0.81 on the tested dataset. 'O' recognition seems to be the most difficult (test F1-measures: 0.90/'N', 0.81/'A' and 0.72/'O') due to high inter-patient variability and similarity with 'N'.

These study results contribute to multidisciplinary areas, focusing on creation of robust and reliable cardiac monitoring systems in order to improve diagnosis, reduce unnecessary time-consuming expert ECG scoring and, consequently, ensure timely and effective treatment.