Enhancing Word-Level Imagined Speech BCI Through Heterogeneous Transfer Learning.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electroencephalogram (EEG)-based Brain-Computer Interfaces (BCI) for Imagined Speech (IS) face challenges in word-level classification accuracy due to complex feature variations.
• Binary classification of IS versus rest yields higher accuracy, suggesting a potential for feature transfer.

Purpose of the Study:

• To introduce Focused Speech Feature Transfer Learning (FSFTL), a novel heterogeneous transfer learning approach.
• To enhance the performance of word-level IS BCI by leveraging features from a high-accuracy binary IS/Rest classification task.

Main Methods:

• Developed FSFTL to utilize a feature extractor trained on binary IS/Rest classification for word-level IS tasks.
• Applied the feature extractor to a public dataset for a five-word IS task, retraining the classifier and fine-tuning the feature extractor.
• Aligned data from different datasets to ensure feature extractor versatility.

Main Results:

• The FSFTL approach demonstrated significant improvements compared to existing EEG models.
• Achieved a 6% increase in mean accuracy across fifteen subjects compared to the backbone strategy.
• Validated the transferability of EEG features for IS across datasets and tasks.

Conclusions:

• FSFTL effectively enhances word-level IS BCI performance by transferring learned features from simpler tasks.
• Highlights the commonality and transferability of EEG features in IS.
• Provides a beneficial strategy for improving the decoding capabilities of word-level IS BCIs.