Enhancing transfer performance across datasets for brain-computer interfaces using a combination of alignment...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Transfer learning (TL) and deep learning (DL) are increasingly used to address variability in brain-computer interfaces (BCIs).
• Existing TL and DL algorithms are often validated on single datasets, assuming identical acquisition conditions, which is unrealistic for cross-dataset generalization.
• This limitation necessitates validation in cross-dataset scenarios to assess true generalization ability.

Purpose of the Study:

• To compare the transfer performance of pre-trained deep learning models with various preprocessing strategies in a cross-dataset setting.
• To evaluate the effectiveness of adaptive batch normalization (AdaBN) in reducing cross-dataset covariate shift.
• To investigate the fusion of manifold embedded knowledge transfer (MEKT) with deep learning models for improved BCI performance.

Main Methods:

• Utilized four public motor imagery datasets, treating each as a source dataset sequentially for cross-dataset validation.
• Trained EEGNet and ShallowConvNet models with four preprocessing strategies: channel normalization, trial normalization, Euclidean alignment, and Riemannian alignment.
• Employed adaptive batch normalization (AdaBN) and compared performance against a manifold embedded knowledge transfer (MEKT) baseline, also exploring MEKT and EEGNet fusion.

Main Results:

• Deep learning models incorporating alignment strategies demonstrated significantly superior transfer performance compared to other preprocessing methods.
• Adaptive batch normalization (AdaBN) effectively improved transfer performance as an unsupervised domain adaptation technique.
• The combination of AdaBN and alignment strategies outperformed MEKT, with EEGNet models achieving the highest generalization across datasets when enhanced by AdaBN and MEKT's domain adaptation.

Conclusions:

• Combining alignment strategies with AdaBN offers a straightforward method to enhance the generalizability of deep learning models in BCIs without requiring fine-tuning.
• This research provides valuable insights for designing transfer neural networks for BCIs, particularly through the separation of source and target batch normalization layers during domain adaptation.