Multimodal active subspace analysis for computing assessment oriented subspaces from neuroimaging data.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomarker Discovery

Background:

• Summarizing high-dimensional neuroimaging data for biomarker discovery requires frameworks that link brain sub-systems to cognitive traits.
• Existing unsupervised methods lack clinical insight, and supervised methods offer limited interpretability of subspaces.

Purpose of the Study:

• To develop a novel computational framework for extracting robust multimodal brain subspaces associated with specific cognitive or biological traits.
• To improve the interpretability and clinical relevance of neuroimaging data analysis for biomarker discovery.

Main Methods:

• Employs active subspace learning on machine learning model gradients.
• Utilizes clustering to identify salient and consistent subspaces related to target variables.
• Validated using a rigorous cross-validation on an Alzheimer's disease dataset.

Main Results:

• Successfully extracts multimodal brain subspaces specific to clinical assessments (e.g., memory, cognitive skills).
• Maintains predictive performance comparable to standard machine learning algorithms.
• Demonstrates consistency of salient active subspace directions across analyses.

Conclusions:

• The framework facilitates biomarker discovery by uncovering Alzheimer's disease-related brain regions within identified subspaces.
• Enables automated identification of structural and functional brain sub-systems characterizing cognitive changes in disorders like Alzheimer's disease.