In this work, we proposed a method for a binary classification in an EEG-based brain computer interface (BCI) with wavelet packet transform and neural networks. For feature extraction, we introduced a new method which combined the slow cortical potentials (SCPs) and the specific energy from the time-frequency domain in beta-band via the wavelet packet transform. A 3-layer perceptron established by back-propagation and the support vector machines (SVMs) were utilized for classification, respectively. We compared the performance in terms of changing the architecture of the net. The accuracy of BP network was found to be best with 4-5-1 architecture reaching an accuracy of 91.47% on test set. Meanwhile, a SVM with Gaussian kernel revealed an accuracy of 91.13% on the test set, showing that multiple types of information have great advantage over other features