Graph Neural Network-Based Diagnosis Prediction.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Electronic Health Records (EHRs) are crucial for predicting patient diagnoses but are high-dimensional, noisy, and temporal.
• Existing recurrent neural network and attention-based methods struggle with data insufficiency and noise.
• Recent knowledge-guided methods improve accuracy by integrating medical ontologies but partially utilize graph information.

Purpose of the Study:

• To propose an end-to-end robust solution, Graph Neural Network-Based Diagnosis Prediction (GNDP), for accurate patient diagnosis prediction.
• To address limitations in existing methods by fully leveraging medical knowledge graphs and global structure information.
• To enhance the modeling of complex EHR data for improved diagnostic accuracy.

Main Methods:

• Constructing sequential patient graphs that integrate EHR history with medical knowledge graph domain information.
• Designing a robust diagnosis prediction model using a spatial-temporal graph convolutional network.
• Employing multiple spatial-temporal graph convolution units to extract features from sequential graph EHR data.

Main Results:

• The GNDP model demonstrated superior performance compared to state-of-the-art methods on two real-world medical datasets.
• The proposed method achieved better utilization of the medical knowledge graph.
• The model successfully generated robust patient representations for accurate diagnosis prediction.

Conclusions:

• GNDP offers a more effective way to utilize medical knowledge graphs for diagnosis prediction.
• The spatial-temporal graph convolutional network approach enhances the accuracy and robustness of EHR data modeling.
• This method represents a significant advancement in leveraging complex health data for improved patient care.