A Brain-Computer Interface (BCI) is a communication tool allowing people to interact with the environment not requiring any peripheral muscle activity to complete the interaction. Brain signals are acquired and processed by a computer in order to identify a particular type of neural processes called event-related potentials (ERPs). ERPs represent the electrical responses recorded from the brain after specific stimulations, called target stimuli. ERP-based BCIs use an oddball paradigm to elicit the ERP components. The user has to focus on the target stimuli which are inserted in a stream of non-target stimuli. The target and non-target stimuli elicit different brain patterns, which can be detected and exploited by the system. Brain responses to target and non-target stimuli can be discriminated solving a binary classification problem. During the training phase, a linear classifier is trained to identify a hyperplane able to separate target and non-target responses. Given the separating hyperplane and considering a single stimulation sequence, the target class is assigned to the stimulus having the highest decision value. Due to the high-dimensional of the feature vector and to the low signal-to-noise-ratio of EEG signals, the detection of brains responses to the target and non-target stimuli is a hard classification problem. In all ERP-based BCI repeated stimulations allow offsetting the low signal-to-noise ratio of electroencephalograms. However, longer stimulation typically means higher accuracy but lower communication rate. Several early stopping methods, that attempt to reduce the number of repetitions preserving the accuracy level, have been introduced in the literature. We propose an effective, easy to implement and classifier independent early stopping method, showing its potential on different datasets recorded on both healthy subjects and amyotrophic lateral sclerosis patients.