Cross-Dataset Variability Problem in EEG Decoding With Deep Learning.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-Computer Interfaces (BCI) face challenges in practical application due to cross-subject variability in neural data.
• Deep learning models show promise for BCI due to superior generalization and feature representation.
• Current deep learning BCI research often lacks validation across multiple datasets, limiting understanding of generalization capabilities.

Purpose of the Study:

• To validate deep learning models for motor imagery (MI) tasks across eight diverse datasets.
• To investigate the impact of cross-dataset variability on the generalization performance of deep learning BCI models.
• To propose and evaluate an effective strategy to mitigate cross-dataset variability in deep learning BCI.

Main Methods:

• Validated deep learning models on eight motor imagery (MI) datasets.
• Introduced an online pre-alignment strategy to align electroencephalography (EEG) distributions across subjects.
• Applied the pre-alignment strategy before model training and inference to reduce inter-subject variability.

Main Results:

• Cross-dataset variability was confirmed to significantly weaken the generalization ability of deep learning BCI models.
• The proposed online pre-alignment strategy effectively reduced the impact of cross-dataset variability.
• Deep learning models integrated with the online pre-alignment strategy demonstrated significantly improved cross-dataset generalization.

Conclusions:

• Deep learning models can be enhanced for BCI applications by addressing cross-subject variability.
• The online pre-alignment strategy offers a viable method to improve BCI model robustness and generalizability.
• This approach enables better cross-dataset performance without requiring additional subject-specific calibration data.