EMG prediction from motor cortical recordings via a nonnegative point-process filter

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Kalman filters are suboptimal for non-Gaussian neural spike train observations.
• Generalized linear models can capture complex neural activity covariates.
• Predicting electromyogram (EMG) signals is crucial for understanding motor control.

Purpose of the Study:

• To develop a constrained point-process filtering mechanism for predicting EMG signals from neural spike recordings.
• To address limitations of Kalman filters in non-Gaussian and non-linear scenarios.
• To model the nonlinear relationship between neural activity and EMG signals.

Main Methods:

• Modeled non-Gaussian neural spike train observations using a generalized linear model.
• Reformulated the Kalman filter within an optimization framework with a nonnegativity constraint.
• Recorded EMG signals from 12 forearm and hand muscles of a behaving monkey during a grip-force task.

Main Results:

• The constrained point-process filter demonstrated improved EMG prediction accuracy with limited training data compared to a Wiener cascade filter.
• Prediction accuracy improvements were observed across various bin sizes and spike-EMG delays.
• For extensive training datasets, the proposed filter's performance was comparable to the Wiener cascade filter.

Conclusions:

• The proposed constrained point-process filter effectively characterizes the nonlinear correspondence between neural activity and EMG signals.
• This method offers a robust approach for EMG signal prediction, particularly when training data is scarce.
• The filter provides a valuable tool for advancing research in neural decoding and motor control.