Supervised nonlinear dimension reduction of functional magnetic resonance imaging data using Sliced Inverse...

Area of Science:

• Neuroimaging
• Machine Learning
• Biostatistics

Background:

• High-dimensional functional magnetic resonance imaging (fMRI) data requires dimension reduction for identifying predictive features.
• Conventional linear methods fail when the relationship between fMRI data and behavioral parameters is nonlinear.

Purpose of the Study:

• To introduce a novel supervised dimension reduction technique, PC-SIR (Principal Component - Sliced Inverse Regression), for high-dimensional fMRI data analysis.
• To address the limitations of existing methods in handling nonlinear relationships and high dimensionality in fMRI data.

Main Methods:

• Proposed PC-SIR, an extension of Sliced Inverse Regression (SIR), incorporating Principal Component Analysis (PCA) to handle high-dimensional data.
• Validated PC-SIR through simulations, comparing its performance against Support Vector Regression (SVR) and Partial Least Square Regression (PLSR).
• Applied PC-SIR to real fMRI data from a pain stimulation experiment involving 32 subjects.

Main Results:

• Simulations demonstrated that PC-SIR achieves more accurate brain activation identification and better prediction than SVR and PLSR.
• Application to real fMRI data showed significantly higher prediction accuracy for pain perception using PC-SIR compared to SVR and PLSR.
• Identified pain-related brain regions using PC-SIR in the fMRI dataset.

Conclusions:

• PC-SIR is a powerful tool for effective dimension reduction in high-dimensional fMRI data, particularly when nonlinear relationships are present.
• PC-SIR offers superior performance over SVR and PLSR for multivariate pattern analysis of fMRI data, enhancing brain region identification and behavioral prediction.
• PC-SIR represents a promising advancement for analyzing complex neuroimaging data in neuroscience research.