Penalized least squares regression methods and applications to neuroimaging

Area of Science:

• Statistics
• Neuroimaging
• Machine Learning

Background:

• Traditional predictor selection methods fail when the number of variables (P) exceeds the number of participants (N).
• High-dimensional datasets, common in neuroimaging, require advanced statistical approaches for accurate analysis.
• Penalized least squares methods like LASSO and Elastic Net are popular but lack uncertainty quantification for selected predictors.

Purpose of the Study:

• To review popular predictor selection methods and explain their limitations in high-dimensional settings.
• To introduce bootstrap-enhanced LASSO (BE-LASSO) and Elastic Net (BE-Enet) for improved variable selection with uncertainty.
• To apply these novel methods to a multimodal neuroimaging dataset for neurocognitive performance assessment.

Main Methods:

• Focus on penalized least squares regression methods: LASSO and Elastic Net.
• Introduce bootstrap enhancements (BE-LASSO, BE-Enet) to quantify uncertainty in variable selection.
• Utilize a multimodal neuroimaging dataset with morphometric, diffusion, and clinical variables where P > N.

Main Results:

• BE-LASSO and BE-Enet successfully perform biomarker selection and dimension reduction in a high-dimensional neuroimaging dataset.
• These methods enable the assessment of neurocognitive performance from complex predictor sets, including interactions.
• The analysis provides a measure of uncertainty for each selected neuroimaging and clinical predictor.

Conclusions:

• Bootstrap-enhanced penalized regression methods (BE-LASSO, BE-Enet) are effective for high-dimensional data analysis.
• These methods offer a significant advantage over traditional approaches by providing uncertainty estimates for selected predictors.
• The study demonstrates a novel statistical framework for integrating neuroimaging and clinical data to predict neurocognitive outcomes.