Multimodal Autism Spectrum Disorder Diagnosis Method Based on DeepGCN.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Multimodal data integration is crucial for diagnosing complex brain diseases like autism spectrum disorder (ASD).
• Existing deep learning methods for ASD identification struggle to efficiently utilize diverse data types.
• Functional magnetic resonance imaging (fMRI) and demographic data offer complementary insights into brain function and individual characteristics.

Purpose of the Study:

• To propose a novel multimodal deep learning framework, WL-DeepGCN, for enhanced autism spectrum disorder (ASD) classification.
• To address the limitations of current deep learning models in integrating heterogeneous data for ASD diagnosis.
• To improve the accuracy and robustness of ASD identification by leveraging both imaging and non-imaging data.

Main Methods:

• Construction of a whole-brain functional connectivity network using fMRI data.
• Development of a weight-learning (WL) network to represent non-imaging data similarity in a latent space for graph edge weight construction.
• Implementation of graph convolutional neural network (GCN) residual connectivity and an EdgeDrop strategy to mitigate information loss and overfitting/oversmoothing issues.

Main Results:

• The proposed WL-DeepGCN model achieved 77.27% accuracy and 0.83 AUC for ASD identification.
• Nested 10-fold cross-validation demonstrated superior performance compared to competitive models.
• The method effectively integrates multimodal data, showing benefits in defining pairwise associations in the latent space.

Conclusions:

• The WL-DeepGCN approach offers a robust and effective multimodal strategy for autism spectrum disorder diagnosis.
• Integrating functional connectivity networks with non-imaging data via a weight-learning graph convolutional network significantly enhances classification performance.
• This study provides a promising direction for leveraging multimodal data in deep learning for neurological disorder identification.