Continuous tracking using deep learning-based decoding for noninvasive brain-computer interface.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Noninvasive brain-computer interfaces (BCIs) using electroencephalography offer potential for interaction without muscle activation.
• Current BCIs have limitations in performance consistency and degrees of freedom, restricting their applications.

Purpose of the Study:

• To investigate the efficacy of deep learning (DL)-based decoders for a complex BCI task: online continuous pursuit (CP).
• To evaluate DL model performance and compare it against traditional BCI decoders.

Main Methods:

• Developed a labeling system for CP data to enable supervised learning.
• Trained DL decoders using two architectures, including a novel PointNet adaptation.
• Evaluated decoder performance across multiple online sessions with 28 participants.

Main Results:

• DL-based models demonstrated performance improvement throughout sessions with increasing training data.
• DL decoders significantly outperformed a traditional BCI decoder by the final session.
• Midsession model updates showed potential benefits, while subject pretraining did not significantly enhance performance.

Conclusions:

• Deep learning decoders are effective for enhancing BCI performance in complex tasks like continuous pursuit.
• Improved BCI performance can broaden the applicability of BCI devices.
• These advancements hold promise for improving the quality of life for individuals with and without motor impairments.