Beyond Feature Importance: Feature Interactions in Predicting Post-Stroke Rigidity with Graph Explainable AI

Graph-based AI models can reveal hidden patterns that traditional methods miss: The study found that graph-based approaches (Graphormer and Graph Attention Networks) outperformed conventional models like logistic regression and XGBoost by explicitly capturing interactions between medical features—such as how stroke severity scores combine with mortality risk factors—rather than treating each factor in isolation.

Explainability matters for clinical decision-making: By using graph explainable AI, researchers could identify not just which factors predict post-stroke rigidity, but how those factors work together. This transparency is crucial for doctors to understand and trust AI recommendations for early intervention in the 43% of stroke patients who develop rigidity.

Better predictions could improve patient outcomes: The graph models achieved an AUROC of 0.75 in predicting post-stroke rigidity, a condition affecting mobility and quality of life. Earlier, more accurate identification could enable timely rehabilitation strategies and personalized treatment plans rather than delayed interventions.

Responsible AI requires moving beyond single-factor analysis: The research demonstrates that responsible AI in healthcare should go beyond identifying individual important features—it should illuminate feature interactions and relationships, providing a more complete and clinically meaningful picture of how medical conditions develop.

Real-world scale validation strengthens AI credibility: Testing on 519,000 stroke hospitalization records from a national dataset (rather than small studies) provides stronger evidence that these graph-based methods could reliably support clinical practice.