Analyzing Data Augmentation Methods for Convolutional Neural Network-based Brain-Computer Interfaces

Brain-computer interfaces (BCI) are systems that use brain signals to communicate with and control devices, with applications ranging over multiple domains. In healthcare, one of the major applications of BCIs is neurorehabilitation. For example, BCIs help stroke patients recover motor abilities by providing sensory feedback based on imagined movement. Convolutional neural networks (CNN) can be used to classify such motor imagery electroencephalogram (EEG) signals and provide this kind of feedback. However, since these signals are usually noisy and can differ significantly over time and among people, it is frequently necessary to collect a large amount of data to train these models. This process can be time-consuming and fatiguing for the user, impairing the quality of neurorehabilitation treatments and other applications. This paper investigates how data augmentation can mitigate this problem by reducing the need for data and increasing feedback accuracy. We analyze five data augmentation methods from the literature on two motor imagery datasets. We apply data augmentation to a few-parameter CNN in varying settings of EEG electrodes, motor imagery tasks, and number of training samples. Our results show that data augmentation can reduce the amount of original data needed, leading to superior accuracy with 33.33 % fewer training samples in some instances. They also show that combining different data augmentation methods can further improve accuracy.