Multi-Modal Multi-Kernel Graph Learning for Autism Prediction and Biomarker Discovery.

Area of Science:

• Computational neuroscience
• Machine learning
• Medical informatics

Background:

• Multi-modal data integration for disease prediction presents challenges due to negative impacts between modalities.
• Existing graph learning methods often rely on static, manually constructed graphs, limiting adaptability.

Purpose of the Study:

• To propose a novel method, Multi-modal Multi-Kernel Graph Learning (MMKGL), for effective multi-modal integration and disease prediction.
• To address the negative impact of modalities during integration and information extraction.
• To identify discriminative brain regions associated with autism.

Main Methods:

• Developed a multi-modal graph embedding module for adaptive graph construction from individual modalities.
• Introduced function and supervision graphs for optimization during multi-graph fusion embedding.
• Employed a multi-kernel graph learning module with convolutional kernels of varying receptive fields to extract heterogeneous information.
• Generated a cross-kernel discovery tensor for disease prediction.

Main Results:

• The proposed MMKGL method demonstrated superior performance on the Autism Brain Imaging Data Exchange (ABIDE) dataset compared to state-of-the-art methods.
• MMKGL successfully identified discriminative brain regions associated with autism.
• The model's findings offer potential guidance for understanding autism pathology.

Conclusions:

• MMKGL offers an effective approach for multi-modal data integration and disease prediction.
• The method's ability to adaptively learn graphs and extract heterogeneous information enhances prediction accuracy.
• The identified brain regions provide valuable insights into autism's underlying mechanisms.