Partial maximum correntropy regression for robust electrocorticography decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Partial Least Square Regression (PLSR) is effective for predicting continuous variables from electrocorticography (ECoG) signals.
• PLSR's susceptibility to noise in brain recordings can degrade performance.
• Robust methods are needed to improve decoding accuracy in noisy ECoG data.

Purpose of the Study:

• To propose a novel robust variant of PLSR for brain-computer interfaces.
• To address the performance deterioration caused by noise in ECoG signal analysis.
• To enhance the robustness and accuracy of decoding tasks using ECoG signals.

Main Methods:

• Introduced Partial Maximum Correntropy Regression (PMCR), a robust implementation of PLSR.
• Utilized the maximum correntropy criterion (MCC) for noise resistance.
• Employed half-quadratic optimization for robust dimensionality reduction and fixed-point optimization for regression coefficients.

Main Results:

• PMCR demonstrated superior prediction results on a synthetic dataset compared to existing methods.
• PMCR effectively extracted valid information in noisy regression scenarios with fewer decomposition factors.
• PMCR achieved superior decoding performance and minimal neurophysiological pattern deterioration on an ECoG dataset.

Conclusions:

• The proposed PMCR method outperforms existing techniques in noisy, inter-correlated, and high-dimensional decoding tasks.
• PMCR effectively alleviates performance degradation caused by noise in ECoG data.
• PMCR enhances the robustness of electrocorticography decoding for brain-computer interfaces.