Graph Artificial Intelligence in Medicine.

Area of Science:

• Clinical Artificial Intelligence (AI)
• Graph Representation Learning
• Machine Learning in Healthcare

Background:

• Clinical datasets contain intricate relationships and structures within diverse data modalities (e.g., patient records, imaging).
• Traditional AI models may struggle to holistically process and learn from these complex, interconnected data structures.

Purpose of the Study:

• To highlight the capabilities of graph representation learning, particularly graph neural networks and transformers, in clinical AI.
• To explore how graph AI can process diverse clinical data by representing entities and modalities as interconnected nodes.
• To discuss the potential and challenges of graph AI in clinical decision-making, focusing on interpretability and human-centered design.

Main Methods:

• Utilizing graph neural networks and graph transformer architectures to model relationships within clinical datasets.
• Representing diverse data modalities and entities as nodes within a graph structure, interconnected by their relationships.
• Leveraging knowledge graphs to enhance model interpretability by aligning AI insights with established medical knowledge.

Main Results:

• Graph AI models can holistically process diverse clinical data, capturing intricate relationships and structures.
• These models facilitate effective model transfer across clinical tasks and patient populations with minimal retraining.
• Graph AI offers opportunities for interpretability through localized transformations and alignment with medical knowledge.

Conclusions:

• Graph AI, through advanced architectures, shows significant promise for analyzing complex clinical data and improving generalization.
• Integrating human-centered design and knowledge graphs is crucial for enhancing the interpretability and clinical utility of AI models.
• Emerging graph AI models with pretraining and interactive feedback loops are paving the way for clinically meaningful predictions and human-AI collaboration.