Feature Selection Based on SVM Significance Maps for Classification of Dementia

Support vector machine significance maps (SVM p-maps) previously showed clusters of significantly different voxels in dementia-related brain regions. We propose a novel feature selection method for classification of dementia based on these p-maps. In our approach, the SVM p-maps are calculated on the training set with a time-efficient analytic approximation. The features that are most significant on the p-map are selected for classification with an SVM classifier. We validated our method using MRI data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), classifying Alzheimer's disease (AD) patients, mild cognitive impairment (MCI) patients who converted to AD within 18 months, MCI patients who did not convert to AD, and cognitively normal controls (CN). The voxel-wise features were based on gray matter morphometry. We compared p-map feature selection to classification without feature selection and feature selection based on t-tests and expert knowledge. Our method obtained in all experiments similar or better performance and robustness than classification without feature selection with a substantially reduced number of features. In conclusion, we proposed a novel and efficient feature selection method with promising results.