Most of the feature extraction methods in existing brain-computer interfaces (BCIs) are based on the dynamic behavior of separate signals, without using the coupling information between different brain regions. In this paper, amplitude and phase coupling measures, quantified by a nonlinear regressive coefficient and phase locking value respectively, were used for feature extraction. The two measures were based on three different coupling methods determined by neurophysiological a priori knowledge, and applied to a small number of electrodes of interest, leading to six feature vectors for classification. Five subjects participated in an online BCI experiment during which they were asked to imagine a movement of either the left or right hand. The electroencephalographic (EEG) recordings from all subjects were analyzed offline. The averaged classification accuracies of the five subjects ranged from 87.4% to 92.9% for the six feature vectors and the best classification accuracies of the six feature vectors ranged between 84.4% and 99.6% for the five subjects. The performance of coupling features was compared with that of the autoregressive (AR) feature. Results indicated that coupling measures are appropriate methods for feature extraction in BCIs. Furthermore, the combination of coupling and AR feature can effectively improve the classification accuracy due to their complementarities.