fMRI activation pattern recognition: A novel application of PCA in Language Network of Pediatric Localization Related Epilepsy

In this study, a novel application of Principal Component Analysis (PCA) is proposed to detect language activation map patterns. These activation patterns were obtained by processing functional Magnetic Resonance Imaging (fMRI) studies on both control and localization related epilepsy (LRE) patients as they performed an auditory word definition task. Most group statistical analyses of fMRI datasets look for "commonality" under the assumption of the homogeneity of the sample. However, inter-subject variance may be expected to increase in large "normal" or otherwise heterogeneous patient groups. In such cases, certain different patterns may share a common feature, comprising of small categorical sub-groups otherwise hidden within the main pooling statistical procedure. These variant patterns may be of importance both in normal and patient groups. fMRI atypical-language patterns can be separated by qualitative visual inspection or by means of Laterality Indices (LI) based on region of interest. LI is a coefficient related to the asymmetry of distribution of activated voxels with respect to the midline and it lacks specific spatial and graphical information. We describe a mathematical and computational method for the automatic discrimination of variant spatial patterns of fMRI activation in a mixed population of control subjects and LRE patients. Unique in this study is the provision of a data-driven mechanism to automatically extract brain activation patterns from a heterogeneous population. This method will lead to automatic self-clustering of the datasets provided by different institutions often with different acquisition parameters.