Causal graph neural networks for healthcare

Area of Science:

• Biomedical informatics
• Artificial intelligence in healthcare
• Causal inference

Background:

• Healthcare AI systems often fail in new settings due to learning spurious correlations instead of causal relationships.
• This leads to performance degradation and perpetuates data biases, hindering reliable deployment.

Purpose of the Study:

• To review causal graph neural networks (CGNNs) for robust healthcare AI.
• To explore CGNNs' potential for creating patient-specific causal digital twins.
• To address challenges and outline future directions for CGNNs in medicine.

Main Methods:

• Utilizing structural causal models and disentangled causal representation learning.
• Applying techniques for interventional prediction and counterfactual reasoning on graphs.
• Integrating graph-based biomedical data with causal inference.

Main Results:

• CGNNs learn invariant causal mechanisms, overcoming limitations of statistical association-based AI.
• Applications demonstrated in psychiatric diagnosis, cancer subtyping, physiological monitoring, and drug recommendations.
• Potential for patient-specific causal digital twins to support in silico clinical experimentation.

Conclusions:

• CGNNs offer a path towards more reliable and equitable healthcare AI.
• Significant challenges remain in computational cost, validation, and preventing 'causal-washing'.
• Future work requires scalable causal discovery, multimodal integration, and clear regulatory pathways for causal AI.