Machine learning methods applied for electroencephalogram present challenges specially due to the limited sample sizes. The main objective of this work is to implement a data augmentation strategy for EEG datasets, which are composed of broad types of information and may be impaired by a limited amount of samples. The method splits subsections of the original data while maintaining key information regarding overall synaptic dynamic, which is then provided as information for training and evaluating predictive models. We observe that it is possible to successfully implement three types of deep learning models with consistently improved accuracy using the presented method for data augmentation, in contrast to simply utilizing the original data. We also investigate the influence of multiple random factors associated with the dataset and obtained results, as well as how such variables may be addressed in the future in order to further enhance the evaluated metric scores.