Graph-based clinical recommender: Predicting specialists procedure orders using graph representation learning.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Health Services Research

Background:

• Access to specialized medical care is limited, causing significant delays in diagnosis and treatment.
• Automated recommender systems can streamline patient referrals and specialist evaluations.

Purpose of the Study:

• To evaluate the efficacy of graph neural network (GNN) models in predicting the need for endocrinology and hematology consultations.
• To compare GNN-based predictions against standard care checklists and existing medical recommendation algorithms.

Main Methods:

• Developed a novel heterogeneous graph neural network model using structured electronic health records.
• Formulated the prediction of subsequent specialist orders as a link prediction problem within a graph framework.
• Trained and assessed models on data from endocrinology and hematology specialty care sites.

Main Results:

• The GNN model demonstrated an 8% improvement in ROC-AUC for endocrinology (0.88) and a 5% improvement for hematology (0.84) compared to prior systems.
• For endocrinology referrals, the GNN recommender achieved higher precision (0.60) and F1-score (0.37) than clinical checklists.
• The GNN model also outperformed checklists for hematology referrals in precision (0.44) and F1-score (0.41).

Conclusions:

• Graph neural network models can significantly enhance the accuracy of digital specialty consultation systems.
• Integrating GNNs into clinical workflows can improve access to specialized medical expertise by leveraging patterns from similar past cases.