Marrying Medical Domain Knowledge With Deep Learning on Electronic Health Records: A Deep Visual Analytics Approach.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Data Visualization

Background:

• Deep learning shows promise in healthcare but faces challenges in interpretability.
• Existing models struggle to integrate complex medical knowledge and offer visual exploration.
• Clinicians need tools to understand and interact with abstract electronic health record (EHR) data and model outputs.

Purpose of the Study:

• Develop an interpretable and accurate risk prediction model for clinical use.
• Create an interactive system to aid EHR data exploration and model interpretation.
• Integrate medical domain knowledge into deep learning models for enhanced prediction.

Main Methods:

• Proposed a domain-knowledge-guided recurrent neural network (DG-RNN) model for clinical risk prediction.
• Incorporated medical domain knowledge via a knowledge graph attention mechanism.
• Developed DG-Viz, an interactive system for EHR data exploration, knowledge graph visualization, and model interpretation.

Main Results:

• The DG-RNN model achieved a 1.5% performance improvement over state-of-the-art methods in heart failure risk prediction.
• A case study confirmed DG-Viz's effectiveness for data exploration and result interpretation by a medical professional.
• The system successfully identified key medical events contributing to clinical risks.

Conclusions:

• DG-Viz integrates deep learning and visual analytics for accurate and interpretable clinical risk prediction.
• The system facilitates better understanding and interaction with EHR data and predictive models.
• This work advances interactive, interpretable, and accurate clinical risk prediction in healthcare.