Feature Selection Based on the SVM Weight Vector for Classification of Dementia

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Diagnostics

Background:

• Computer-aided diagnosis of dementia can be enhanced through effective feature selection.
• Support Vector Machines (SVMs) are utilized for classification, with SVM weights offering insights into feature relevance via significance maps (p-maps).
• Previous studies identified relevant brain regions using p-maps, but their application in feature selection was unexplored.

Purpose of the Study:

• Introduce and evaluate novel feature selection methods for dementia diagnosis based on p-maps.
• Compare p-map-based methods against SVM weight vectors, t-statistics, and expert knowledge.
• Assess the efficacy of these methods in classifying Alzheimer's Disease (AD), mild cognitive impairment (MCI), and cognitively normal (CN) individuals.

Main Methods:

• Developed two p-map feature selection methods: a direct (filter) approach and an iterative (wrapper) approach.
• Utilized MRI data from the Alzheimer's Disease Neuroimaging Initiative (ADNI).
• Extracted features from gray matter morphometry for voxel-wise analysis.

Main Results:

• P-map based feature selection methods slightly outperformed direct SVM weight vector methods.
• The wrapper approach demonstrated superior performance compared to the filter approach.
• Recursive feature elimination using p-maps significantly improved the Area Under the Curve (AUC) for Alzheimer's Disease vs. Cognitively Normal classification from 90.3% to 92.0% by selecting 1.5%-3% of features.

Conclusions:

• P-map based feature selection methods show promise for improving dementia diagnosis accuracy.
• These methods offer a more effective way to estimate the relevance of individual features compared to traditional approaches.
• While performance improvements were modest, p-map methods consistently yielded better results, highlighting their potential in neuroimaging analysis.