[Paper Review] Data augmentation for learning predictive models on EEG: a systematic comparison

Cédric Rommel, Joseph Paillard+2 more|arXiv (Cornell University)|Jun 29, 2022

EEG and Brain-Computer Interfaces38 references96 citations

TL;DR

This paper systematically evaluates 13 data augmentation techniques for EEG-based predictive modeling across sleep stage classification and motor imagery BCI tasks. It demonstrates that appropriate augmentation can boost accuracy by up to 45% in low-data regimes, with performance gains varying significantly by task and data type, highlighting the need for task-specific augmentation strategies rather than one-size-fits-all approaches.

ABSTRACT

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

Motivation & Objective

• To address the challenge of small, high-variability EEG datasets that hinder deep learning generalization.
• To evaluate the effectiveness of diverse data augmentation techniques across multiple EEG tasks and models.
• To identify which augmentations yield the greatest performance gains under low-data conditions.
• To determine whether augmentation benefits are consistent across different data classes and training set sizes.
• To provide a unified benchmark for EEG data augmentation that informs future model development.

Proposed method

• The study employs a unified experimental protocol across two EEG tasks: sleep stage classification (SleepPhysionet dataset) and motor imagery BCI (BNCI-2014-1 dataset).
• It evaluates 13 augmentation methods across three domains: time (e.g., GaussianNoise, TimeReverse), frequency (e.g., FTSurrogate, FrequencyShift), and spatial (e.g., ChannelsSymmetry, SignFlip).
• Each augmentation is applied with varying magnitudes and evaluated using cross-validation with stratified splits to preserve class distribution.
• Models are trained using a deep CNN architecture with consistent hyperparameters across all experiments to isolate the effect of augmentation.
• Performance is measured via accuracy, with comparisons made across different training set sizes and data classes to assess robustness and generalization.
• The analysis includes ablation studies to assess the impact of augmentation magnitude and class-specific effects.

Experimental results

Research questions

• RQ1 Which data augmentation techniques yield the most significant accuracy improvements in low-data regimes for EEG classification?
• RQ2 How do the benefits of data augmentation vary across different EEG tasks (e.g., sleep staging vs. motor imagery BCI)?
• RQ3 Does the effectiveness of an augmentation depend on the magnitude of the transformation applied?
• RQ4 How do augmentation effects differ across data classes, particularly in imbalanced datasets?
• RQ5 Is there a single optimal augmentation strategy, or does performance depend on the specific task and data characteristics?

Key findings

• Data augmentation can improve model accuracy by up to 45% in low-data regimes compared to models trained without augmentation.
• The most effective augmentations vary significantly between tasks: FTSurrogate and FrequencyShift were most beneficial for sleep stage classification, while TimeReverse and SignFlip outperformed others for motor imagery BCI.
• Performance gains from augmentation are highly dependent on the magnitude of the transformation, with moderate perturbations yielding the best results.
• Augmentations that preserve spectral or temporal structure (e.g., FTSurrogate, FrequencyShift) were more effective than those that disrupt signal integrity.
• No single augmentation strategy is universally optimal; the best method depends on the specific EEG task, data modality, and class distribution.
• The study confirms that EEG classification tasks benefit substantially from appropriate data augmentation, especially when training data is limited.