Data augmentation for learning predictive models on EEG: a systematic\n comparison

Abstract

Objective: The use of deep learning for electroencephalography (EEG)\nclassification tasks has been rapidly growing in the last years, yet its\napplication has been limited by the relatively small size of EEG datasets. Data\naugmentation, which consists in artificially increasing the size of the dataset\nduring training, can be employed to alleviate this problem. While a few\naugmentation transformations for EEG data have been proposed in the literature,\ntheir positive impact on performance is often evaluated on a single dataset and\ncompared to one or two competing augmentation methods. This work proposes to\nbetter validate the existing data augmentation approaches through a unified and\nexhaustive analysis. Approach: We compare quantitatively 13 different\naugmentations with two different predictive tasks, datasets and models, using\nthree different types of experiments. Main results: We demonstrate that\nemploying the adequate data augmentations can bring up to 45% accuracy\nimprovements in low data regimes compared to the same model trained without any\naugmentation. Our experiments also show that there is no single best\naugmentation strategy, as the good augmentations differ on each task.\nSignificance: Our results highlight the best data augmentations to consider for\nsleep stage classification and motor imagery brain-computer interfaces. More\nbroadly, it demonstrates that EEG classification tasks benefit from adequate\ndata augmentation\n