Disease prediction using graph convolutional networks: Application to Autism Spectrum Disorder and Alzheimer's...

Area of Science:

• Neuroscience
• Medical Imaging Analysis
• Machine Learning

Background:

• Graphs model interactions in populations, useful for medical data with imaging and non-imaging features.
• Existing graph methods for disease prediction often overlook individual features or fail to model interactions effectively.

Purpose of the Study:

• To evaluate a generic framework leveraging both imaging and non-imaging information for brain analysis in large populations.
• To assess the framework's performance in disease prediction tasks, specifically Autism Spectrum Disorder and Alzheimer's disease conversion.

Main Methods:

• Utilized Graph Convolutional Networks (GCNs) to represent populations as sparse graphs.
• Integrated imaging features as node attributes and phenotypic information as edge weights.
• Evaluated framework components and compared performance against baselines on ABIDE and ADNI datasets.

Main Results:

• Achieved 70.4% classification accuracy for Autism Spectrum Disorder prediction on the ABIDE dataset.
• Reached 80.0% classification accuracy for Alzheimer's disease conversion prediction on the ADNI dataset.
• Demonstrated improved performance over state-of-the-art methods on both datasets.

Conclusions:

• The proposed GCN-based framework effectively integrates diverse data types for improved disease prediction.
• This approach offers a robust method for brain analysis in large populations, outperforming existing methods.