Supplementary Table S1: Time-series ECG data of arrhythmic patients and normal controls smoothened by Savitzky-Golay filter.

Full-length ECG time series data of 48 arrhythmic patients has been collected from popular MIT-BIH Arrhythmia Database [1-3]. The data sets include ECG time series data of 25 men aged between 32 to 89 years and of 22 women aged between 23 to 89 years. On the other hand, the normal data (healthy person) has been collected from the MIT-BIH Normal Sinus Rhythm Database [4] consisting of 18 long-term ECG signals of subjects having no significant arrhythmia. The subjects include 5 men and 13 women, aged between 26 to 45 and 20 to 50 respectively. The signal used here for the analysis is a modified limb lead II (MLII), obtained by placing the electrodes on the chest of the patients. We did not include the records of two patients with patient IDs 102 and 104 from MIT-BIH Arrhythmia Database as the required MLII data were not available due to surgical dressings on the above-mentioned patients. On the other hand, we also consider the data from ECG1 mode which are ECG signals (Normal Sinus Rhythm Database) relating to healthy persons and these data sets are regarded as complementary to MLII data of the arrhythmia database. Here, total no. of data points of each of the disease data series is 21600 with frequency 360.01 per sec whereas the same for the normal data series is 7680 with frequency 128 per sec. Therefore, each of the data series is recorded for 60 sec time duration. A filter is utilized for processing of signals in order to selectively isolate a particular frequency or range of frequencies from an assortment of multiple frequencies in a signal. The choice of appropriate filter for processing of the system generated signals requires maximum noise reduction with minimal signal distortion [5]. One of the best filters for noise clearing of biomedical data, including ECG signals, seems to be Savitzky–Golay (SG) filter [6]. The fundamental principle of SG filter is to consider (2n + 1) equidistant points taking n = 0 as a centre to represent a polynomial of degree p (where p < (2n + 1)). A set of points is to be fitted to some curve. For this purpose, SG filter computes the value of the least square polynomial (or its derivative) at a point, i = 0, over the decided frame range. This filter applies the method of linear least squares for data smoothing, which helps to maintain the original shape of the signal. The SG filter generally requires pre-determined values of order and frame depending on the frequency and length of the data. Usually, trial and error method or prior experience is required to decide the satisfactory values of parameters. Here, the values of “Frame” for diseased and normal data were assumed to be 37 and 13 respectively and the “Order” of the filter were 3 for each type of data sets. The ‘Supplementary Table SI’ contains the filtered data sets for both types of disease and normal subject.

本数据集从知名的MIT-BIH心律失常数据库（MIT-BIH Arrhythmia Database）[1-3]中采集了48例心律失常患者的全长度心电（Electrocardiogram, ECG）时间序列数据。数据集包含25名年龄介于32至89岁的男性，以及22名年龄介于23至89岁的女性的心电时间序列数据。
另一方面，健康受试者的正常心电数据采自MIT-BIH正常窦性心律数据库（MIT-BIH Normal Sinus Rhythm Database）[4]，该库包含18份无显著心律失常受试者的长期心电信号。其中受试者为5名年龄介于26至45岁的男性，以及13名年龄介于20至50岁的女性。
本研究采用的分析信号为改良肢体导联II（Modified Limb Lead II, MLII），通过在患者胸部放置电极采集得到。我们未纳入MIT-BIH心律失常数据库中患者ID为102与104的两份记录，原因是上述患者因手术敷料覆盖，无法获取所需的MLII数据。
此外，本研究还纳入了来自正常窦性心律数据库的ECG1模式数据，即健康受试者的心电信号，该类数据集可作为心律失常数据库MLII数据的补充。
各类疾病数据序列的数据点总数均为21600，采样频率为360.01 Hz；正常数据序列的数据点总数为7680，采样频率为128 Hz。因此，每份数据序列的记录时长均为60秒。
本研究采用滤波器对信号进行处理，以从信号中的多种频率成分里选择性分离出特定频率或频率范围。针对系统生成信号的滤波处理，需在尽可能降低噪声的同时最小化信号失真[5]。在包括心电信号在内的生物医学数据降噪处理中，萨维茨基-戈莱（Savitzky–Golay, SG）滤波器被认为是最优选择之一[6]。
SG滤波器的基本原理为：以(2n+1)个等距点为基础，以n=0为中心拟合次数为p的多项式（其中p < 2n+1），即通过最小二乘法将一组点拟合至某一曲线。具体而言，SG滤波器会在选定的窗口范围内，计算某点i=0处的最小二乘多项式（或其导数）值。该滤波器通过线性最小二乘法实现数据平滑，有助于保留信号的原始形态。
SG滤波器的阶数与窗口大小通常需根据数据的频率与长度预先确定，一般需通过试错法或凭借先验经验确定合适的参数取值。本研究中，疾病数据与正常数据对应的窗口大小（Frame）分别设为37与13，两类数据集的滤波器阶数（Order）均设为3。
‘补充表SI’包含了两类受试者（疾病组与正常组）经过滤波处理后的数据集。