Interpreting support vector machine models for multivariate group wise analysis in neuroimaging.

Area of Science:

• Neuroimaging
• Machine Learning
• Biostatistics

Background:

• Machine learning, particularly support vector machines (SVMs), shows potential for classifying high-dimensional neuroimaging data.
• Identifying specific imaging features driving SVM decisions is crucial for clinical utility and understanding disease mechanisms.
• Existing statistical inference methods for SVMs often overlook the critical SVM margin.

Purpose of the Study:

• To develop and validate a novel statistic for statistical inference in SVM-based neuroimaging classification.
• To address the limitation of existing methods by explicitly incorporating the SVM margin.
• To improve the interpretability and reliability of machine learning models in neuroimaging research.

Main Methods:

• Development of a new statistic that accounts for the SVM margin.
• Asymptotic normality analysis of the null distributions for the proposed statistic.
• Experimental comparison with traditional weight-based permutation tests.

Main Results:

• The proposed statistic explicitly accounts for the SVM margin, a key theoretical component.
• The null distributions associated with the new statistic are shown to be asymptotically normal.
• Experiments demonstrate the new statistic is less conservative and more specific than weight-based permutation tests.
• The statistic effectively identifies multivariate patterns used by SVMs in neuroimaging classification.

Conclusions:

• The novel statistic offers a more robust approach to statistical inference in SVM-based neuroimaging.
• This method enhances the interpretability of machine learning models by revealing key contributing imaging patterns.
• Improved understanding of multivariate patterns aids in critical evaluation of findings and disease mechanism exploration.