Lightweight beat score map method for electrocardiogram-based arrhythmia classification

We recently investigated beat score map (BSM)-based methods for electrocardiogram (ECG)-based arrhythmia classification. Although BSM-based methods show impressive performance, they are somewhat resource-intensive owing to the arrangement of beat score vectors generated from 1D ECG sequences with zero-padding across time points. To address this issue, we propose a lightweight BSM (Lw-BSM) method that significantly reduces the size of the original BSM while capturing the characteristics of beat arrangement patterns as does the original BSM. Specifically, two types of Lw-BSMs are generated without zero-padding and evaluated for multiclass arrhythmia prediction. Experimental results on two public datasets, MIT-BIH and SPH, demonstrate that arrhythmia classification using Lw-BSM images is quite comparable to that using the original BSM images as an input to CNN-based classification models. At the same time, the image size can be reduced significantly. Moreover, it is observed that this approach is almost insensitive to the selection of the R-peak detection algorithm, showing stable performance across different R-peak algorithms.