Low-Rank Representation for Multi-center Autism Spectrum Disorder Identification

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Multi-center data is crucial for understanding autism spectrum disorder (ASD) pathology.
• Existing machine learning methods often fail to account for data heterogeneity across centers, impacting diagnostic performance.
• Data heterogeneity arises from variations in scanning parameters and subject populations.

Purpose of the Study:

• To propose a novel multi-center low-rank representation learning (MCLRR) method for improved ASD diagnosis.
• To effectively address and mitigate data heterogeneity in multi-center datasets.
• To develop a robust representation learning approach for cross-center ASD diagnosis.

Main Methods:

• A multi-center low-rank representation learning (MCLRR) framework was developed.
• Domain-specific projections were learned to transform source domains into an intermediate representation space.
• Projection matrices were decomposed into shared and sparse unique parts to reduce heterogeneity.
• A k-nearest neighbor (KNN) classifier was used for disease classification based on the learned representation.

Main Results:

• The proposed MCLRR method demonstrated superior performance in ASD diagnosis compared to existing methods.
• The approach effectively handled data heterogeneity from multiple centers.
• Classification accuracy was significantly improved on the ABIDE database.

Conclusions:

• The MCLRR method offers an effective solution for leveraging multi-center data in ASD diagnosis.
• Addressing data heterogeneity is critical for enhancing the reliability of machine learning models in clinical applications.
• This approach holds promise for improving diagnostic accuracy and understanding of ASD.