Brain-computer interfaces (BCI) based on motor imagery tasks (MI) have been established as a promising solution for restoring communication and control of people with motor disabilities. Physically impaired people may perform different motor imagery tasks which could be recorded in a non-invasive way using electroencephalography (EEG). However, the success of the MI-BCI systems depends on the reliable processing of the EEG signals and the adequate selection of the features used to characterize the brain activity signals for effective classification of MI activity and translation into corresponding actions. The multilayer perceptron (MLP) has been the neural network most widely used for classification in BCI technologies. The fact that MLP is a universal approximator makes this classifier sensitive to overtraining, especially with such noisy, non-linear, and non-stationary data as EEG. Traditional training techniques, as well as more recent ones, have mainly focused on the machine-learning aspects of BCI training. As a novel alternative for BCI training, this work proposes an incremental training process. Preliminary results with a non-disabled individual demonstrate that the proposed method has been able to improve the BCI training performance in comparison with the cross-validation technique. Best results showed that the incremental training proposal allowed an increase of the performance by at least 10% in terms of classification compared to a conventional cross-validation technique, which indicates the potential application for classification models of BCI’s systems. © Springer Nature Switzerland AG 2019.