Two-step paretial least square regression classifiers in brain-state decoding using functional magnetic resonance...

Area of Science:

• Neuroimaging
• Machine Learning
• Brain-Computer Interfaces

Background:

• Multivariate analysis is crucial for decoding brain states from fMRI data.
• Partial Least Squares Regression (PLSR) is commonly used for feature selection but less so as a direct classifier.
• The performance of PLSR as a classifier in fMRI brain-state decoding remains largely unexplored.

Purpose of the Study:

• To propose and evaluate novel two-step PLSR classifiers for enhanced brain-state decoding from fMRI data.
• To investigate the efficacy of using PLSR and sparse PLSR (SPLSR) for both feature selection and classification.
• To compare the performance of the proposed PLSR classifiers against traditional methods like GLM and SVM.

Main Methods:

• Development of two-step PLSR classifiers integrating PLSR/SPLSR for feature selection and PLSR for classification.
• Validation using simulated and real fMRI datasets.
• Comparative analysis with PLSR-GLM and SVM-PLSR/SPLSR/GLM classifiers.

Main Results:

• The proposed PLSR classifiers, utilizing PLSR/SPLSR for feature selection, outperformed PLSR-GLM and SVM-based classifiers in most scenarios.
• PLSR with SPLSR for feature selection demonstrated the highest performance among all tested methods.
• PLSR exhibited greater sensitivity in identifying task-specific voxels compared to GLM.

Conclusions:

• The performance of PLSR classifiers for fMRI brain-state decoding can be substantially enhanced by integrating feature selection methods like SPLSR and PLSR.
• The proposed two-step PLSR approach offers a promising avenue for improving brain-state decoding accuracy in neuroimaging research.