In event-related potentials based brain-computer interfaces, the responses evoked by a well defined stimuli sequence are usually averaged to overcome the limitations caused by the intrinsic poor EEG signal-to-noise ratio. This, however, implies that the time necessary to detect the brain signals increases and then that the communication rate can be dramatically reduced. A common approach is then at first to estimate an optimal fixed number of responses to be averaged on a calibration data set and then to use this number on the online/testing dataset. In contrast to this strategy, several early stopping methods have been successfully proposed, aiming at dynamically stopping the stimulation sequence when a certain condition is met. We propose an efficient and easy to implement early stopping method that outperforms the ones proposed in the literature, showing its effectiveness on several publicly available datasets recorded from either healthy subjects or amyotrophic lateral sclerosis patients.