Mdpg: a novel multi-disease diagnosis prediction method based on patient knowledge graphs.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems

Background:

• Electronic Health Record (EHR) data presents challenges for diagnosis prediction due to its time-series, sparse, and multi-noise nature.
• Existing methods often use Recurrent Neural Networks (RNNs) and medical knowledge bases but overlook local spatial characteristics and spatial-temporal correlations.
• Accurate diagnosis prediction is crucial for enhancing healthcare efficiency and clinical decision support.

Purpose of the Study:

• To propose MDPG, a novel diagnosis prediction model leveraging patient knowledge graphs.
• To effectively capture local spatial structure, temporal characteristics, and spatial-temporal correlations within EHR data.
• To improve the accuracy and efficiency of predicting future patient health states.

Main Methods:

• Representing electronic visit records as patient-centered temporal knowledge graphs.
• Developing spatial graph convolution, temporal self-attention, and spatial-temporal synchronous graph convolution blocks.
• Utilizing multi-label classification for future state prediction and evaluating with visit-level precision@k and code-level accuracy@k metrics.

Main Results:

• The proposed MDPG model demonstrated superior performance compared to all baseline models.
• Comprehensive experiments on two real-world datasets validated the effectiveness of MDPG.
• MDPG successfully captured spatial, temporal, and spatial-temporal correlations in EHR data.

Conclusions:

• MDPG offers a significant advancement in diagnosis prediction using patient knowledge graphs.
• The model's ability to capture complex data correlations leads to improved prediction accuracy.
• This approach enhances clinical decision support by providing more reliable future patient state predictions.