Objective The rehabilitation technology based on brain-computer interface (BCI) has become a crucial issue for the patient with motor dysfunction to achieve movement. The key technique of BCI is to quickly and accurately identify the EEG mode which is associated with motor Imagery. An adaptive algorithm of classification based on wavelet packet transform and kernel extreme learning machine (ELM) algorithm is proposed according to the characteristic of EEG such as Non-stationary and individualized differences and so on to enhance the classification accuracy of EEG. Method As the existence of the cross-banding of wavelet packet, the average energy of the best wavelet packet basis which is extract adaptively using distance criterion form the feature vector, and the kernel ELM algorithm is applied for classification. BCI competition data are used for the classification of the proposed method. The classification accuracy of different algorithms is simulated and analyzed. Results Simulation results demonstrate that the average classification accuracy is achieved to 97.6%and outperforms state-of-the-art algorithms such as ELM, back propagation (BP) and support vector machine (SVM) in the aspects of training time and classification accuracy. Conclusions The proposed method produces a high classification accuracy and is suitable for EEGclassification application.