A novel feature extraction method, which uses subband features calculated from higher order statistics, is proposed for ECG beat classification. Five levels of discrete wavelet transformation (DWT) are applied to decompose the signal into six subband signals with different frequency distribution. Higher order statistics proceeds to calculate valuable features from the three midband signals. Three RR interval-related features are added to build a feature vector of 30 features. The feature extraction paradigm cooperates with the typical feedforward backpropagation neural network (FFBNN) to discriminate seven ECG beat types. Two signal selection profiles are also proposed for different experimental design. Between them, the 2nd profile which selected different beat types from each record was determined to be more appropriate and was used as a test bench for comparing with the other methods. The proposed method demonstrates a promising accuracy of 97.53%. Comparing to the method which used higher order statistics solely to the original signal, the proposed method elevates the sensitivities of most beat types. Especially in the three beat types: RBBB, VEB, and VFW, more than 4% elevations in sensitivity are observed. This study demonstrates the effectiveness of applying higher order statistics to subband signals for characterizing ECG beats for computer-aided diagnosis of heart diseases.