IKDP: Implicit Knowledge Enhanced Disease Prediction via heterogeneous admission sequence graphs.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computational Biology

Background:

• Deep learning models for electronic health records (EHRs) struggle with complex disease relationships and patient admission trajectories.
• Existing knowledge graph-based methods are limited by incomplete knowledge and overlook implicit patient data like similarities and latent correlations.
• Discarding single-admission patients leads to loss of valuable clinical insights.

Purpose of the Study:

• To develop an advanced model for disease prediction using EHR data.
• To address limitations in current EHR modeling by incorporating implicit patient information.
• To enhance the representation of complex disease relationships and patient admission trajectories.

Main Methods:

• Introduced the Implicit Knowledge Enhanced Disease Prediction model (IKDP) utilizing heterogeneous admission sequence graphs (SeqGs).
• Integrated an auxiliary pre-training strategy with end-to-end optimization for multi-dimensional patient data processing.
• Computed inter-patient similarities as complementary knowledge and constructed SeqGs to capture disease dependencies and health status evolution.

Main Results:

• The IKDP model effectively harnesses implicit knowledge from comprehensive patient admission data.
• SeqGs capture complex disease dependencies and the dynamic evolution of patient health status.
• Critical paths from SeqGs, similar patient analysis, and historical records elucidate prediction reasoning.

Conclusions:

• IKDP offers a novel approach to EHR modeling by leveraging implicit knowledge and heterogeneous sequence graphs.
• The model improves the representation of patient data, including those with single admissions.
• This method enhances disease prediction accuracy and provides interpretable insights into clinical reasoning.