Care coordination and patient safety outcome: a graph-based approach

Abstract Predicting postoperative adverse events and managing the associated risk factors is crucial for patient safety. Care coordination, also known as provider team interactions, significantly impacts outcomes, yet few studies have explored this link and applied it to risk prediction. To address this, Medical Heterogeneous Graphs for Patient Safety analysis (MedHG-PS), a novel graph-based framework that simultaneously models complex relationships among patient characteristics, provider interactions, and patient transfer records was proposed in this study. Evaluated on a real-world dataset with 102,768 patients from the University of Florida Health Integrated Data Repository, MedHG-PS outperforms state-of-the-art methods, achieving an AUC above 0.90 and up to a 20% improvement in recall for three major postoperative outcomes—prolonged length of stay (PLOS), 30-day mortality, and 90-day mortality. By using meta-path analysis (MPA), SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), MedHG-PS identifies key predictive features, such as patient transfers highly influencing PLOS, whereas provider interactions affect mortality risks. This study highlights how care coordination can be modeled at scale using EHRs and can affect patient care safety outcomes—an important aspect of an automated and rapid-learning health system.