Hierarchical Mixture-of-Experts Framework for Whole-Brain Imaging and Whole-Genome Analysis of Multiple Cognitive...

Area of Science:

• Neuroscience
• Genetics
• Artificial Intelligence

Background:

• Assessing brain health and cognitive function requires understanding complex biological mechanisms.
• Neuroimaging and genetic data offer complementary insights but present challenges due to high dimensionality and heterogeneity.
• Integrating these multimodal data for accurate cognitive function prediction remains a significant hurdle.

Purpose of the Study:

• To develop an effective framework for joint task-guided feature selection and multimodal fusion.
• To enhance the prediction of cognitive scores by integrating whole-brain imaging and whole-genome data.
• To address the challenges of high-dimensional and heterogeneous multimodal data in cognitive neuroscience.

Main Methods:

• Proposed IGMoE, a hierarchical mixture-of-experts framework.
• Utilized modality-specific sparse mixture-of-experts networks (SMoE) for feature extraction and selection from imaging and genetic data.
• Employed contrastive pretraining for multimodal representation alignment and a fusion mixture-of-experts network (FMoE) for dynamic integration.

Main Results:

• IGMoE demonstrated superior performance in predicting cognitive scores compared to existing methods.
• The framework effectively performed multi-task whole-brain and whole-genome analysis using both unimodal and multimodal data.
• Validation was conducted on UK Biobank, ADNI, and PPMI datasets for various cognitive tasks.

Conclusions:

• IGMoE provides an effective solution for integrating multimodal neuroimaging and genetic data for cognitive function analysis.
• The proposed framework advances the field of multi-task whole-brain and whole-genome analysis.
• This approach holds promise for improving brain health assessments and intervention strategies.