Mini-batch sampling strategies for data augmentation in MI-BCI decoding based on deep learning

CATALINA MARÍA GALVÁN; SPIES, RUBEN DANIEL; DIEGO H. MILONE; VICTORIA PETERSON

Bruselas

Congreso; 10th International Brain-Computer Interface Meeting 2023; 2023

BCI Society

Introduction: The typical approach to implement data augmentation strategies for deep learning (DL) models is to generate new training samples by suitable modifications of the original data, and then train the model via mini-batch stochastic gradient descent [1]. In the context of DL for brain computer interfaces (BCIs), using synthetically generated electroencephalography (EEG) trials can be considered as a better approach for data augmentation. However, while most of the current DL training strategies employ random data sampling, for BCIs applications this might not be the most appropriate option due to data scarcity. Moreover, synthesized data can be thought of as samples from a domain different from the original one and this fact must be taken into account when sampling for the construction of mini-batches. Here we study different sample selection techniques for data augmentation in DL-based BCI decoding. Material, Methods and Results: Experiments were performed in a cross-session scenario (train: session 1, and test: session 2) in the right vs. left hand motor imagery (MI) OpenBMI dataset [2]. Each real session included 100 trials per class. FBCNet [3] was used as decoding model. Data augmentation samples consisted of 160 simulated MI-EEG trials (extending [4]), with the inclusion of subject-specific aperiodic and periodic information. Four different mini-batch sample selection techniques were evaluated: 1) standard random sampling (RS), where the ratio of real/augmented data and right/left hand cases change randomly between mini-batches; 2) domain-aware sampling (DAS), in which all mini-batches have the same ratio of real/augmented samples, 3) stratified domain-aware sampling (SDAS), in which mini-batches are stratified by output class and the ratio of real/augmented samples is also fixed. For the three methods, the batch size was set to 16 and for DAS and SDAS the ratio real/augmented was 14/2. No differences were found by adding more augmented samples in each mini-batch. The results showed that the SDAS achieves the highest average accuracy value (66.23% ± 2.88%), followed by DAS (65.90% ± 2.75%) and RS (65.28%± 2.45%). The baseline accuracy (i.e., without data augmentation) was 64.86% ± 2.68%. According to the Nemenyi test [5], SDAS is significantly better (p-value