Sub-100 μW Multispectral Riemannian Classification for EEG-Based Brain-Machine Interfaces.

Area of Science:

• Neuroscience
• Computer Science
• Electrical Engineering

Background:

• Wearable brain-machine interfaces (BMIs) require on-device processing for privacy and usability.
• Existing embedded solutions face challenges in balancing accuracy and energy efficiency.

Purpose of the Study:

• To investigate the accuracy-cost trade-off for embedded motor imagery (MI) brain-machine interface (BMI) solutions.
• To develop and evaluate an energy-efficient, near-sensor classification model for wearable MI-BMIs.

Main Methods:

• Implemented a multispectral Riemannian classifier and optimized it per subject.
• Quantized the model to mixed-precision representations to reduce resource requirements.
• Deployed the model on a low-power microcontroller unit (MCU) and analyzed energy consumption and classification speed.

Main Results:

• Achieved 75.1% accuracy on a 4-class MI task, improved to 76.4% with subject-specific tuning.
• Quantization resulted in minimal accuracy loss (1%-1.4%), outperforming state-of-the-art embedded convolutional neural networks by up to 4.1%.
• The model operates within a 198 μJ energy budget per classification (16.9 ms) and achieves 85 μW continuous operation.

Conclusions:

• The developed embedded MI-BMI system achieves state-of-the-art accuracy-energy trade-off for near-sensor classification.
• The solution offers practical insights for creating efficient and accurate wearable BMI devices.
• The approach ensures privacy, user comfort, and long-term usability for controlling external machines via thought.