Feature Selection for Dementia Classification Using Support Vector Machine

Feature selection is of great importance in medical image classification especially neuroimaging classification for determining the most relevant features that will aid in accurate diagnosis of neuropsychological diseases. This paper presents a comparison of feature selection algorithms based on Support Vector Machine (SVM). To achieve robust performance and optimal selection of parameters involved in feature selection, and classification, prior knowledge is embedded to generate multiple versions of training and testing sets for parameter optimization. The integrated feature extraction and selection method is applied to a Structural Magnetic Resonance image based Alzheimer's dementia (AD) study with four different sets of non-demented and demented subjects. Cross-validation results of our study clearly indicate that the algorithm SVM-RFE trained with prior knowledge achieves 98% accuracy with Radial Basis Function (RBF) kernel and can improve performance of the classifier. This novel method of inculcating prior knowledge in SVM-RFE method which is tested in 4 different sets of datasets reveals that RBF kernel is found to outperform other kernels with a mean sensitivity of 97%, and thereby aids in quick and efficient classification of dementia.