Parameter optimization of 3D convolutional neural network for dry-EEG motor imagery brain-machine interface

Area of Science:

• Biomedical Engineering
• Neuroscience
• Artificial Intelligence

Background:

• Brain-machine interfaces (BMI) offer solutions for individuals with care needs by alleviating behavioral restrictions and reducing caregiver burden.
• Edge computing for BMI enhances data privacy and system ubiquity but faces hardware limitations for complex deep learning models.
• Optimizing deep learning models is crucial for implementing efficient and accurate BMI systems on resource-constrained edge devices.

Purpose of the Study:

• To develop a high-accuracy, low-latency motor imagery (MI) classification system for brain-machine interfaces (BMI) suitable for edge deployment.
• To optimize deep learning models and measurement conditions to minimize computational cost, power consumption, and response latency while maintaining classification accuracy.
• To investigate the efficacy of a 3-dimensional convolutional neural network (3D CNN) for improved MI classification and adapt it for practical use with dry electrodes.

Main Methods:

• Implemented a deep learning approach focusing on motor imagery (MI) classification using electroencephalography (EEG) data.
• Optimized deep learning model parameters, including kernel size, number, and layer structure, alongside MI measurement conditions.
• Utilized a 3-dimensional convolutional neural network (3D CNN) and adapted the model for use with dry electrodes, optimizing for fewer electrodes, reduced recall time, and lower sampling rates.

Main Results:

• The proposed 3D CNN model achieved high classification accuracy for MI with reduced computational cost, power consumption, and latency, suitable for edge processing.
• Compared to EEGNet, the 3D CNN demonstrated significant reductions in parameters (75.9%), multiply-accumulates (16.3%), and memory footprint (12.5%).
• The system maintained classification accuracy under practical conditions, including eight electrodes, a 3.5-second sample window, and a 125 Hz sampling rate with dry EEG.

Conclusions:

• The optimized 3D CNN provides an effective solution for edge-based BMI systems, enhancing motor imagery classification accuracy and efficiency.
• This approach can significantly improve the quality of life for individuals requiring care and reduce the workload for care providers.
• The developed BMI system demonstrates a practical and privacy-preserving method for assistive technology in nursing care settings.