Benchmarking Deep Learning Methods for Arrhythmia Detection

Automatic arrhythmia detection methods are a very significant area of computational ECG analysis. This field has been researched for a long time, however, there are various challenges still faced. Some of the main flaws in current ECG-based arrhythmia classification research are limited variety of datasets used and varying experimental setups, which makes it difficult to directly compare different methods. Most often, a method is evaluated on a specific dataset and task (set of arrhythmia classes). By placing these methods under unified evaluation setup (one umbrella), we can apply (evaluate) them on a wider range of datasets and tasks than they were originally proposed for. To address these challenges, in this paper, we perform benchmarking of some of the most significant deep-learning based methods for arrhythmia detection. These methods are compared on four datasets, considering the most significant state-of-the-art arrhythmia classification tasks. Included are the data from the CinC2017 and CPSC2018 challenges, as well as two recently published large-scale ECG arrhythmia datasets: the PTB-XL and the Shaoxing Hospital Database. The analyses cover a wide range of both morphological and rhythmic arrhythmias, all while focusing on methods suitable for single-lead analysis. In addition, the classification performance on 12-lead data and single-lead data is compared and discussed.