Supervised Phenotype Discovery From Multimodal Brain Imaging.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biostatistics

Background:

• Image-derived phenotypes (IDPs) from brain imaging hold potential for neuroscience and clinical research.
• Current IDP discovery methods often lack integration with non-imaging derived phenotypes (nIDPs).

Purpose of the Study:

• To introduce SuperBigFLICA, a semi-supervised, multimodal, multi-task fusion approach for enhanced IDP discovery.
• To integrate information from multiple imaging modalities and nIDPs simultaneously.

Main Methods:

• Developed SuperBigFLICA, a computationally efficient semi-supervised fusion approach.
• Applied SuperBigFLICA to the UK Biobank brain imaging dataset (approx. 40,000 subjects, 47 modalities, >17,000 nIDPs).

Main Results:

• SuperBigFLICA significantly enhanced nIDP prediction accuracy (up to 46% improvement) compared to existing methods.
• The approach learned generic imaging features capable of predicting novel nIDPs.
• Demonstrated robustness across various prediction tasks and biological meaningfulness in predicting health and cognitive outcomes.

Conclusions:

• SuperBigFLICA offers a powerful and efficient method for data-driven IDP discovery.
• Integrating multimodal imaging and non-imaging data advances neuroscientific and clinical insights.
• The method facilitates the discovery of IDPs linked to health outcomes and cognitive functions.