In this paper, we discuss a subject-based feature extraction method using the fuzzy wavelet packet in brain–computer interfaces (BCIs). The method includes the following three steps: (1) original electroencephalogram (EEG) signals are decomposed with the wavelet packet transform (WPT), which forms many wavelet packet bases; (2) for each subject and each EEG channel, the best basis algorithm based on a fuzzy set criterion is used to find the best-adapted basis for that particular subject and channel; and (3) subband energies included in the best basis form effective features, which are used to discriminate three types of motor imagery tasks. The proposed method is compared with the previous wavelet packet method and the results show that it outperforms the previous one.