Identifying autism spectrum disorder from multi-modal data with privacy-preserving

Area of Science:

• Artificial Intelligence in Medicine
• Neuroscience
• Medical Informatics

Background:

• Deep learning for precision medical diagnosis requires large datasets, but data privacy hinders multi-institutional data aggregation.
• Current automatic diagnosis for autism spectrum disorder (ASD) using multimodal data shows suboptimal performance.
• Existing methods struggle with privacy preservation and effective fusion of heterogeneous medical data.

Purpose of the Study:

• To propose a privacy-preserving deep learning framework for improved ASD diagnosis using multimodal data.
• To address the challenge of aggregating sensitive medical data across institutions.
• To enhance the performance of ASD identification through advanced feature fusion techniques.

Main Methods:

• Developed a federated learning framework (FedHNN) integrating multimodal feature fusion and hypergraph neural networks.
• Employed federated learning for distributed model training without direct data sharing, ensuring privacy.
• Utilized a hypergraph fusion strategy to capture inter-modal complementarity and correlations for ASD diagnosis.

Main Results:

• The proposed FedHNN model demonstrated superior performance compared to individual local models.
• FedHNN outperformed other deep learning models in ASD identification tasks.
• The framework effectively leveraged multi-site data for improved diagnostic accuracy while preserving privacy.

Conclusions:

• Federated learning combined with hypergraph neural networks offers a robust solution for privacy-preserving, multi-institutional medical data analysis.
• The FedHNN framework significantly improves the performance of autism spectrum disorder diagnosis.
• This approach facilitates the development of scalable and accurate deep learning models for precision medicine.