Graph-based Fusion Modeling and Explanation forDisease Trajectory Prediction

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Personalized clinical event prediction is crucial for managing diseases like COVID-19.
• Existing methods often lack the ability to capture complex patient similarities beyond basic demographics.
• Integrating diverse data sources like imaging and electronic health records (EHR) presents a significant challenge.

Purpose of the Study:

• To develop a relational graph model for personalized clinical event prediction in hospitalized COVID-19 patients.
• To effectively fuse heterogeneous data, including chest X-rays (CXRs) and non-imaging EHR data, for enhanced predictive modeling.
• To capture complex clinical patterns and patient similarities without manual feature selection.

Main Methods:

• Constructing a relational graph where nodes represent patient snapshots and weighted edges encode clinical similarity.
• Utilizing CXRs as node features and non-imaging EHR data for edge formation.
• Developing a model that learns from individual patient data and patterns within clinically similar patient cohorts.

Main Results:

• The graph-based approach effectively incorporates clinical similarity, leading to improved personalized event prediction.
• Visualization studies demonstrated that the model leverages the 'neighborhood' of a node, focusing on clinically similar patients.
• The model showed a tendency to identify and utilize shared suggestive clinical features among connected patients.
• The proposed method demonstrated good generalization by adapting the edge formation process to external cohorts.

Conclusions:

• Relational graph construction offers a powerful framework for integrating heterogeneous clinical data for predictive modeling.
• The model's ability to learn from clinically similar patients enhances the accuracy and personalization of event prediction.
• This approach avoids manual feature engineering and adapts to new patient populations, showing significant potential for clinical applications.