Robust and accurate decoding of hand kinematics from entire spiking activity using deep learning

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-machine interfaces (BMIs) aim to restore motor function for individuals with neurological disorders.
• Clinical translation of intracortical BMIs faces challenges in robustness and decoding accuracy.
• Current methods often struggle with long-term performance and signal variability.

Purpose of the Study:

• To enhance the robustness and decoding accuracy of intracortical brain-machine interfaces.
• To introduce a novel input signal, entire spiking activity (ESA), for improved BMI performance.
• To evaluate the efficacy of ESA coupled with a deep learning decoder.

Main Methods:

• Proposed using entire spiking activity (ESA), extracted via a simple, automated technique, as the input signal.
• Employed a deep learning decoding algorithm utilizing a quasi-recurrent neural network (QRNN) architecture.
• Evaluated the ESA-driven QRNN decoder on chronically recorded neural signals from non-human primates performing various tasks.

Main Results:

• The ESA-driven QRNN decoder consistently achieved higher decoding performance than previously reported methods.
• High decoding performance was maintained even when spikes were removed or the number of channels was varied.
• The method demonstrated sustained accuracy with reduced training data, indicating improved robustness.

Conclusions:

• The proposed method demonstrates exceptionally high decoding accuracy and chronic robustness for brain-machine interfaces.
• This advancement addresses key unresolved challenges in the clinical translation of BMIs.
• The findings suggest a promising direction for restoring motor function in individuals with neurological impairments.