Feature selection and extraction in sequence labeling for arrhythmia detection

Automated Electrocardiogram (ECG)-based arrhythmia detection methods replace traditional, manual arrhythmia detection reducing the requirement for trained medical staff. Traditionally, ECG-based arrhythmia detection is performed via QRS complex detection followed by feature extraction, based on hand-crafted features, such as RR-intervals, Fast Fourier Transform-based features, wavelet analysis, higher order statistics and Hermite features. After the features are extracted, the ECG segments are classified into pre-defined categories. This study investigates the value of the feature extraction and selection methods for ECG-based arrhythmia detection. That is, with the emerging trend of deep learning methods which are capable of automatic feature extraction and selection, the research question addressed in this paper is if good classification performance can be obtained by feeding the raw ECG sequence directly into robust classifiers or handcrafted feature extraction/selection is necessary. Classification performance across a range of state-of-the-art classification methods indicates that feeding raw signals into the convolution neural network-based classifiers usually leads to the best performance but at the expense of high inference time