Preprocessing: Although it may cause the loss of real useful ECG signals, it is a critical step for improving classification accuracy before measuring the amplitudes to perform baseline drift correction for ECG data. One way is to use the integer coefficient digital filter and the 10-point moving average filter [36]. Another way is to remove the baseline wander by median filter [4], and remove the high frequency noise by symN wavelet, such as sym10 [37].
• R-wave detection: There are many QRS detection methods including So-and-Chan QRS detection algorithm [6] and Hilbert transform with automatic threshold [38].
• Feature extraction: We only adopt the continuous RR intervals (RRI_25-RRI+5o) as attributes of each sample.