REW-NPLS Regression for BCI Online Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current direct neural interfaces face challenges in accurately estimating movement trajectories from electrophysiological signals, particularly in calibrating predictive models for controlling machines like exoskeletons or computers, due to limitations in signal processing and model updating methods.
Innovation Solution
The implementation of Recursive Exponentially Weighted N-way Partial Least Squares (REW-NPLS) regression for online calibration and model updating, which uses tensors to project signals onto latent variables, and introduces penalization terms to enhance sparsity and robustness, allowing for efficient prediction of movement trajectories.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional PLS regression is used for calibration, then the predictive model can be built, but the model updating efficiency and accuracy deteriorate due to lack of online adaptation capability
Solution Approach 1:
The patent transforms the static PLS regression model into a dynamic online calibration system by implementing recursive model updating. The predictive model parameters are continuously adapted using incoming neural signal data streams, allowing the system to maintain high accuracy while efficiently processing data in real-time through incremental updates rather than complete re-calibration.
Solution Approach 2:
The patent performs preliminary dimensionality reduction by projecting neural signal data onto a lower-dimensional latent variable space before performing regression analysis. This pre-processing step reduces computational complexity and enables faster online model updating while preserving the essential information needed for accurate trajectory prediction.
2Measurement precision
If more electrodes are used to improve signal quality, then the measurement accuracy improves, but the device complexity and data processing burden increase
Solution Approach 1:
The patent extracts only the most relevant features from the high-dimensional neural signal data by projecting onto a reduced latent variable space. This extraction process identifies and retains the critical signal components while discarding redundant information, thereby maintaining measurement precision with reduced computational complexity and fewer processing resources.
Solution Approach 2:
The patent transforms the data from high-dimensional electrode space into a lower-dimensional latent variable space through projection. This dimensionality change preserves the essential signal characteristics needed for accurate trajectory estimation while significantly reducing the computational burden and effective complexity of processing signals from multiple electrodes.
3Measurement precision
If online calibration is performed continuously, then the model accuracy is maintained, but the computational energy consumption increases
Solution Approach 1:
The patent performs partial calibration updates by only recalculating the necessary model parameters based on new incoming data, rather than performing complete re-calibration. This partial action approach maintains model accuracy by updating only the essential components that have changed, thereby significantly reducing computational energy consumption while preserving predictive performance.
Data Source
Figure 1
AI summary
The invention relates to an online calibration method for a direct neural interface (BCI) using a penalized multi-way regression method REW-NPLS to update the predictive model of the interface. The hyperparameter values involved in the penalization are determined using a reinforcement learning method during the recursive validation step of the REW-NPLS algorithm.