Using beat score maps with successive segmentation for ECG classification without R-peak detection

As a primary indicator for cardiovascular diseases, the electrocardiogram (ECG) is commonly used for arrhythmia classification. Many related studies emphasize that the R-peaks of the ECG signal are essential for extracting features or signal segmentation. Thus, the chosen R-peak detection algorithm affects classification performance. Furthermore, the lack of distinct R-peaks in arrhythmias like ventricular flutter makes these rhythms difficult to identify, regardless of the detection algorithm. Therefore, this study proposes a novel ECG rhythm classification framework that does not depend on R-peak detection. First, the n number of beat segments is acquired by sliding a window over the ECG signal. A scalogram is then produced from each segment and fed into a pre-trained beat classifier to generate n beat score vectors. These vectors are concatenated chronologically to establish an n-beat score map (n-BSM), which serves as input for our rhythm classification model. The n-BSM of a rhythm conveys information regarding its constituent n beats by sequentially arranging their characteristics, each captured by a score distribution over various beat types. Experimental results from the MIT-BIH arrhythmia database (MITDB) demonstrate that the proposed method improves overall performance in classifying ten ECG rhythms. Moreover, we achieved a 97.56% F1 score for a ventricular flutter rhythm lacking distinct R-peaks. We also utilized the MIT-BIH Malignant Ventricular Ectopy (VFDB) and the Chapman-Shaoxing 12-lead ECG databases (SPH) to verify the proposed method's robustness and generalizability. The average accuracy for different rhythms was 99.28% and 88.83%, respectively.