In nonhuman studies, it is reported that spatiotemporal patterns emerging in spontaneous brain activities reflect past experiences embedded in neural circuits. In human studies, however, spatiotemporal patterns in spontaneous brain activities have not been examined well. This is because extracting spatiotemporal patterns is difficult owing to their unknown onsets. Here, we propose a method to extract spatiotemporal patterns that appear frequently from spontaneous brain activity data, such as MEG and fMRI. This method iteratively searches for spatiotemporal patterns and their onsets, with the aim of minimizing the residual error between the observed and reconstructed data. We tested the performance of this method using real MEG data. From continuous MEG data during a visual stimulation task, this method extracted a spatiotemporal pattern resembling a visual evoked response without using stimulus onset information. This result confirms the applicability of the method for real MEG data. By applying this method, it will be possible to examine how past experiences are engraved in our brains from spontaneous human brain activities.