A selective overview of feature screening methods with applications to neuroimaging data

Area of Science:

• Computational Statistics
• Statistical Learning
• Image Data Mining

Background:

• Neuroimaging studies often involve regression models to associate imaging features with clinical outcomes.
• The number of imaging features (e.g., voxel-level predictors) frequently exceeds the number of subjects, posing a challenge for traditional methods.
• Classical variable selection methods struggle with the ultrahigh-dimensional nature of neuroimaging data.

Purpose of the Study:

• To provide a selective review of recent developments in ultrahigh-dimensional variable screening.
• To focus on the practical performance of these screening methods in neuroimaging analysis.
• To evaluate methods considering complex spatial correlations and high dimensionality.

Main Methods:

• Extensive simulation studies were conducted to compare different variable screening methods.
• Performance was assessed based on selection accuracy and computational costs.
• Analyses included resting-state functional magnetic resonance imaging (fMRI) data from the Autism Brain Imaging Data Exchange study.

Main Results:

• Simulation studies compared the selection accuracy and computational efficiency of various ultrahigh-dimensional variable screening techniques.
• The practical performance of these methods was evaluated on real-world neuroimaging datasets.
• The review synthesizes findings on the effectiveness of different screening approaches for high-dimensional neuroimaging data.

Conclusions:

• Ultrahigh-dimensional variable screening is essential for effective feature selection in neuroimaging.
• The review provides insights into the performance of recent screening methods for complex, high-dimensional data.
• Findings are relevant for advancing computational statistics and image data mining in neuroscience.