COGNITIVE STATE PREDICTION VIA TWO-DIMENSIONAL FEATURE VECTOR

Abstract

Predicting human cognitive tasks from their corresponding functional Magnetic Resonance Imaging (fMRI) data is very challenging. The difficulty of this prediction problem can be summarized in two points first: the size of the dataset is very small due to the small number of subjects (i.e., patients) who can contribute to these research-based experiments, second: the size of feature vector is very large compared to the few number of the samples that can be used to derived a prediction model. One possible way to overcome these obstacles is to develop a feature generation methodology that can result a small-sized and descriptive feature vector that may improve the overall performance of the cognitive state classification problem. Motivated by these considerations, we proposes a novel feature generation methodology termed voxel weightbased (VW) features that can represent the voxels intensity when a subject is performing a certain cognitive task. This feature generation technique can project the high-dimensional feature vector into a two-dimensional feature domain. After generating the VW-based feature set, a logistic regression model (LRM) is utilized to distinguish between two cognitive states that correspond to two distinct tasks (whether a subject is viewing a picture or a sentence). To demonstrate the efficacy of the proposed feature generation scheme, a benchmark fMRI dataset is utilized to assess the performance of the LRM under the proposed features.