Fast BCIs: Leveraging Dual-Scale Time Windows with Test-Time Adaptation to Enhance Accuracy

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) require rapid outputs for practical use.
• Deep learning in BCIs faces challenges with decoding accuracy at short time windows (TW).

Purpose of the Study:

• To develop a novel strategy for fast and accurate BCIs using deep learning.
• To enhance decoding performance in short TWs by integrating long TW information.

Main Methods:

• Introduced a dual-scale time window (DTW) strategy combined with test-time adaptation (TTA).
• Employed two neural networks: a Main Network (MainNet) for short TW decoding and an Auxiliary Network (AuxNet) for long TW pseudo-label generation.
• Utilized AuxNet to update MainNet during testing, improving short TW accuracy.

Main Results:

• DTW-TTA achieved high accuracy and information transfer rates (ITR) across diverse BCI paradigms (MI, SSVEP, ERP) and signal-to-noise ratios (SNRs).
• Demonstrated superior performance compared to state-of-the-art baselines on multiple datasets.
• Specifically, achieved 74.64% accuracy on BCI-IV 2b (MI) and 91.86% on Benchmark (SSVEP) at a 0.5s TW.

Conclusions:

• DTW-TTA effectively stabilizes decoding in short time windows for BCIs.
• The proposed method can be seamlessly integrated into deep learning-based BCI systems.
• This strategy offers a promising approach for achieving both speed and accuracy in BCIs.