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

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Patient Safety Research

Background:

• Predicting postoperative adverse events is critical for patient safety.
• Care coordination (provider team interactions) impacts patient outcomes but is understudied in risk prediction.
• Existing methods lack comprehensive modeling of patient characteristics, provider interactions, and transfer records.

Purpose of the Study:

• To propose Medical Heterogeneous Graphs for Patient Safety analysis (MedHG-PS), a novel graph-based framework.
• To simultaneously model complex relationships among patient data, provider interactions, and patient transfer records.
• To enhance the prediction of postoperative adverse events and identify key risk factors.

Main Methods:

• Developed MedHG-PS, a graph-based framework utilizing heterogeneous graphs.
• Evaluated MedHG-PS on a large-scale real-world dataset (102,768 patients).
• Employed meta-path analysis (MPA), SHapley Additive exPlanations (SHAP), and Local Interpretable Model-agnostic Explanations (LIME) for feature identification.

Main Results:

• MedHG-PS achieved an AUC above 0.90, outperforming state-of-the-art methods.
• Demonstrated up to a 20% improvement in recall for prolonged length of stay (PLOS), 30-day mortality, and 90-day mortality.
• Identified patient transfers as key predictors for PLOS and provider interactions as significant for mortality risks.

Conclusions:

• MedHG-PS effectively models care coordination using Electronic Health Records (EHRs) for patient safety analysis.
• The framework provides insights into how care coordination impacts patient safety outcomes.
• Highlights the potential for automated, rapid-learning health systems to improve patient care.