Neuroscience-Inspired Deep Learning Brain-Machine Interface Decoder.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-machine interfaces (BMIs) decode neural activity for device control.
• Current monolithic decoders struggle with generalizability across different movement directions.

Purpose of the Study:

• To develop a novel, neuroscience-inspired decoder for enhanced BMI performance.
• To improve the generalizability and robustness of motor intention decoding.

Main Methods:

• Proposed a Single-Direction Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) decoder with parallel branches.
• Modeled extension and flexion dynamics separately using spatial-temporal feature extraction from neural spike data.
• Tested the decoder using invasive recordings from a macaque during a 2D center-out reaching task.

Main Results:

• The proposed decoder achieved performance comparable to conventional CNN-LSTM on trained tasks.
• Significantly outperformed CNN-LSTM and linear regression baselines in cross-target generalization.
• Captured physiologically meaningful co-contraction patterns, offering deeper insights into motor control.

Conclusions:

• Neuroscience-inspired modular decoding enhances BMI robustness and adaptability.
• This approach offers a promising pathway for advanced BMI applications in prosthetics and rehabilitation.
• The modular design improves cross-task generalizability in brain-machine interfaces.