Disease Prediction via Graph Neural Networks.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Electronic Medical Records (EMRs) enable disease prediction, but current machine learning models struggle with data scarcity for rare diseases.
• Existing methods often fail to predict diseases for new patients lacking historical EMR data.

Purpose of the Study:

• To develop an innovative Graph Neural Network (GNN) model for accurate disease prediction, addressing limitations of existing approaches.
• To enhance prediction accuracy for both common and rare diseases, even with limited patient data.

Main Methods:

• Constructed a medical concept graph and patient record graph using external knowledge bases and EMRs.
• Utilized Graph Neural Networks (GNNs) to learn representative node embeddings for patients, diseases, and symptoms.
• Developed a neural graph encoder to aggregate information and inductively infer embeddings for new patients.

Main Results:

• The proposed GNN model demonstrated state-of-the-art performance on a real-world EMR dataset.
• Successfully augmented insufficient EMR data using external knowledge bases.
• Achieved accurate disease prediction for both general and rare diseases, including for patients without prior EMR history.

Conclusions:

• The novel GNN-based approach significantly improves disease prediction accuracy, particularly for rare diseases.
• The model's ability to handle new patients enhances its real-world applicability in clinical settings.
• Integrating external medical knowledge with EMR data is crucial for robust disease prediction systems.