A Systematic Evaluation of Euclidean Alignment with Deep Learning for EEG Decoding

TL;DR

This study systematically evaluates Euclidean Alignment (EA) as a lightweight domain adaptation step for deep learning-based EEG decoding in BCI, focusing on cross-subject transfer and ensemble strategies. By comparing EA with Riemannian Alignment and testing across shared and individual models on two MI EEG datasets, the authors show that EA provides consistent target-subject gains (about 4.33% in online-like conditions) and dramatically reduces training convergence time (over 70%), while enabling effective transfer and improved ensemble performance. Fine-tuning EA-backed shared models offers limited or no additional gains, whereas EA enhances transferability and allows stronger majority-voting ensembles among individual models, albeit with some trade-offs in speed. Overall, the results support adopting EA as a standard preprocessing step for cross-subject DL EEG decoding, contributing to more efficient, scalable BCI systems, with clear avenues for future hyperparameter and ensemble-method improvements.

Abstract

Electroencephalography (EEG) signals are frequently used for various Brain-Computer Interface (BCI) tasks. While Deep Learning (DL) techniques have shown promising results, they are hindered by the substantial data requirements. By leveraging data from multiple subjects, transfer learning enables more effective training of DL models. A technique that is gaining popularity is Euclidean Alignment (EA) due to its ease of use, low computational complexity, and compatibility with Deep Learning models. However, few studies evaluate its impact on the training performance of shared and individual DL models. In this work, we systematically evaluate the effect of EA combined with DL for decoding BCI signals. We used EA to train shared models with data from multiple subjects and evaluated its transferability to new subjects. Our experimental results show that it improves decoding in the target subject by 4.33% and decreases convergence time by more than 70%. We also trained individual models for each subject to use as a majority-voting ensemble classifier. In this scenario, using EA improved the 3-model ensemble accuracy by 3.7%. However, when compared to the shared model with EA, the ensemble accuracy was 3.62% lower.

Figures (10)

• Figure 1: Euclidean Alignment steps illustrated. The star represents the origin of the space.
• Figure 2: Shared, individual, and majority-voting classifier models. (a) We use data from all source subjects to train a single shared model, with and without EA. (b) We use data from each source subject to train a separate individual model, with and without EA. (c) We use calibration data from the target subject to select the best source models and combine them in a majority-voting classifier.
• Figure 3: Prediction accuracy using ShallowNet, DeepNet, and EEGNet shared models without alignment, and with EA and RA, in the online and offline scenarios. Left: BNCI2014 dataset, right: Schirrmeister2017 dataset.
• Figure 4: Validation accuracy and loss during training iterations using EEGNet and BCI2014 datasets, with shaded areas depicting the standard deviations over all trained models, one per target subject.
• Figure 5: Significance matrix comparing all shared pipelines, with colored cells indicating a significant difference. The algorithms in the x-axis are the baselines, and the values shown correspond to the standardized mean difference between pipelines in the y-axis and the respective baselines and the p-values.