Graph-Based Deep Learning for Medical Diagnosis and Analysis: Past, Present and Future

Area of Science:

• Biomedical Informatics
• Machine Learning
• Computational Biology

Background:

• Machine learning, particularly deep learning, is increasingly vital for analyzing complex healthcare data.
• Traditional methods struggle with irregular, non-grid-like physiological recordings, limiting their application.
• Graph neural networks (GNNs) offer a powerful alternative by modeling biological systems as interconnected graphs.

Purpose of the Study:

• To provide a comprehensive survey of graph neural network architectures.
• To review the diverse applications of GNNs in healthcare data analysis.
• To identify limitations and future research avenues in GNNs for healthcare.

Main Methods:

• Systematic review of graph neural network methodologies.
• Categorization of GNN applications based on healthcare data domains: functional connectivity, anatomical structure, and electrical-based analysis.
• Analysis of existing GNN techniques and their inherent limitations.

Main Results:

• GNNs effectively analyze irregular physiological data by leveraging relationships between nodes (e.g., temporal, anatomical).
• Applications span functional connectivity, anatomical structure analysis, and electrical-based healthcare assessments.
• Identified limitations include scalability and interpretability challenges.

Conclusions:

• GNNs represent a significant advancement for healthcare data analysis, overcoming limitations of grid-based methods.
• Further research is needed to address current limitations and unlock the full potential of GNNs in clinical applications.
• The survey provides a foundational overview for researchers and practitioners in the field.