Comparative Analysis of Neural Decoding Algorithms for Brain-Machine Interfaces.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Accurate neural decoding is crucial for brain-machine interfaces (BMIs).
• A lack of systematic comparison exists for current decoding algorithms.
• Optimizing algorithm combinations is key for advancing BMI technology.

Purpose of the Study:

• To conduct a large-scale comparative study of state-of-the-art algorithms for motor neural decoding.
• To identify the optimal combination of signal processing, feature extraction, and classification methods.
• To evaluate computational and memory requirements for real-time embedded computing.

Main Methods:

• Evaluated three signal processing methods: artifact subspace reconstruction, surface Laplacian filtering, data normalization.
• Assessed four feature extractors: common spatial patterns, independent component analysis, short-time Fourier transform, no feature extraction.
• Compared four classifiers: support vector machine, linear discriminant analysis, convolutional neural networks, long short-term memory networks, using a large-scale EEG dataset and optimizing for individual subjects (672 experiments).

Main Results:

• Identified optimal algorithm combinations for motor neural decoding based on classification accuracy.
• Compared computational and memory storage requirements for each combination.
• Provided insights into algorithm performance for real-time BMI applications.

Conclusions:

• The systematic comparison offers valuable guidance for selecting and designing neural decoding algorithms.
• Findings will inform the development of next-generation brain-machine interfaces.
• Optimal algorithm selection can enhance BMI performance and efficiency.