Most brain computer interfaces (BCIs) that classify right/left wrist bending movements or their imageries measure and learn the electroencephalogram (EEG) rhythmic modulation involved in sensorimotor activities. However, the issue is that the BCIs do not perform well when the brain activities that do not exist in the learn process occur outside the sensorimotor area. To resolve the issue, we propose the BCI that adopts the reconstructed signal power in the activated region in which the subject performs the movements as features for a linear classifier. The activated region for each subject is identified using functional magnetic resonance imaging (fMRI). We adopt a linearly constrained prewhitening spatial filter for reconstruction of signal power, which is suitable for reconstructing event-related desynchronization/synchronization signals and for existing highly correlated signals. The result is that the proposed BCI that adopts an EEG spatial filter with linear constraints of fMRI activated regions obtains higher accuracy than previous methods.