Optimizing Motor Intention Detection With Deep Learning: Towards Management of Intraoperative Awareness

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) detection from electroencephalographic (EEG) signals is crucial for Brain-Computer Interfaces (BCIs).
• Deep learning techniques offer potential for improved MI detection accuracy.
• Intraoperative awareness monitoring during general anesthesia is a key application area.

Purpose of the Study:

• To investigate deep learning architectures, specifically EEGNet, for enhanced MI detection from raw EEG signals.
• To evaluate the impact of functional electrical stimulation and electrode configurations on MI detection performance.
• To optimize MI detection for potential applications in monitoring intraoperative awareness.

Main Methods:

• Exploration of various EEGNet architectures for motor imagery detection.
• Comparison of EEGNet against state-of-the-art BCI classifiers (Riemannian geometry, LDA) and other deep learning models (DCN, SCN).
• Acquisition of EEG data from 22 participants performing MI with and without median nerve stimulation.

Main Results:

• The proposed EEGNet achieved 83.2% accuracy and 19.0% FPR with 6 electrodes and an extended 4-38 Hz band during median nerve stimulation.
• Higher electrode counts (128) yielded improved accuracy (94.5%) and lower FPR (6.1%).
• A configuration with 13 electrodes resulted in 88.0% accuracy and 12.9% FPR.

Conclusions:

• An extended EEG frequency band and a modified EEGNet deep neural network significantly increase MI detection accuracy.
• Effective MI detection is achievable with as few as 6 electrodes, including frontal channels.
• This research advances the development of BCI systems utilizing MI detection from EEG signals.