Kernel-based feature relevance analysis for ECG beat classification

Abstract

The analysis of Electrocardiogram (ECG) records for arrhythmia classification favors the developing of aid diagnosis systems. However, current devices provide large amounts of data being necessary the development of signal processing methodologies to reveal relevant information. Here, a kernel-based feature relevance analysis approach is introduced to highlight discriminative attributes in ECG-based arrhythmia classification tasks. For such purpose, morphological and spectral-based features are extracted from each provided heartbeat. Then, a linear mapping is learned by using a Kernel Centered Alignment-based scheme to highlight themost relevant features when estimating nonlinear dependencies among samples. The proposed approach is performed as a cascade classification scheme to avoid biased results due to unbalance issue of the studied phenomenon. The results yield a performance rate of 86.52% (sensitivity), 97.57% (specificity), and 92.57% (accuracy) in a well-known database, which validate the reliability of the proposed algorithm in comparison to the state-of-art.