The Merit of Principal Component Analysis in fMRI Language Pattern Recognition for Pediatric Epilepsy

摘要

Atypical language activation pattern analysis is of significant clinical relevance in neuroscience research, especially when surgical interventions are deemed necessary. Epilepsy patient populations provide a means for validating these methods because of known heterogeneity of language dominance. Florida International University (FIU), in collaboration with 13 worldwide health care institutions, has established a multisite repository for language reorganization analysis on normal and pediatric epilepsy fMRI data. Quantitative region of interest (ROI) analysis, Laterality Index (LI) calculation, and visual rating are common methods for determining language dominance. Limitations of subjective ROI analysis with priori assumption or subjective visual rating motivate us to seek a data-driven method. Here we propose a new configuration and application of the Principal Component Analysis for fMRI language activation pattern recognition among a heterogeneous population. The top eigenvectors are proposed to objectively automate the recognition of ROI among fMRI datasets. 122 subjects' fMRI activation maps were processed, visually rated by clinical investigators. ROI identified through the PCA-based method generally encompass Broca's and Wernicke's areas. fMRI datasets masked by the ROI were applied as input to the proposed PCA method. Different numbers of top eigenvectors were examined in comparison to their spatial distributions of LI and their respective visual ratings. These PCA-based brain activation distributions suggest a potential of using eigenvectors to separate and classify fMRI language activation patterns.