Multi-level dynamic fusion and temporal role-aware network for diagnosis prediction.

Area of Science:

• Medical Informatics
• Computational Biology
• Artificial Intelligence in Healthcare

Background:

• Electronic health records (EHR) analysis for diagnosis prediction faces challenges in modeling disease associations and progression.
• Current methods often overlook the clinical significance of disease combinations and dynamic changes.

Purpose of the Study:

• To propose a novel network, the Dynamic Fusion and Role-aware Network (DFR-Net), for enhanced diagnosis prediction.
• To capture the dynamic evolution and progression patterns of diseases within EHR data.

Main Methods:

• Constructing a feature-disease similarity matrix to fuse diagnostic codes with clinical monitoring indicators.
• Defining dynamic disease roles (persistent, evolving, episodic) and using matrixed gating units to model stage-shifting patterns.
• Adaptively fusing diagnostic dependencies with clinical feature relevance for dynamic disease progression analysis.

Main Results:

• The DFR-Net effectively captures associations between clinical monitoring indicators and disease progression.
• The model successfully models stage-shifting patterns and disease evolution mechanisms.
• Experiments on the MIMIC-III dataset show superior performance compared to state-of-the-art diagnosis prediction models.

Conclusions:

• The DFR-Net significantly improves diagnosis prediction accuracy by integrating clinical indicators and disease codes.
• The proposed approach offers a robust framework for analyzing complex disease progression dynamics in EHR.
• This model provides a decision basis for understanding and predicting disease evolution.