Facilitating the Development of Deep Learning Models with Visual Analytics for Electronic Health Records.

Area of Science:

• Medical Informatics
• Data Science in Healthcare
• Clinical Pathway Analysis

Background:

• Electronic health records (EHRs) are crucial for early diagnosis and treatment planning.
• Analyzing large and complex EHR datasets efficiently presents significant challenges.
• Existing methods may lack the speed and interpretability required for iterative data-intensive research.

Purpose of the Study:

• To enhance the efficiency of applying data-intensive technologies to complex EHR data.
• To develop a system for verifying results in big EHR data analysis.
• To facilitate the exploration of clinical pathways and hypothesis testing in cardiology.

Main Methods:

• Utilized sequence mining to identify patient pathways from EHR data.
• Employed interpretable deep learning models for prediction and analysis.
• Integrated visualization tools, including Sankey diagrams and heat maps, for temporal and quantitative illustration.
• Applied the system to predict unplanned cardiac surgery in heart disease patients using MIMIC-III data.

Main Results:

• Sequence mining successfully identified clinically relevant patient pathways.
• The system facilitated the selection of patient groups based on identified pathways.
• Visualization tools provided clear temporal and quantitative insights into patient journeys and variable importance.
• The approach aided in simplifying patient group labeling for deep learning models.

Conclusions:

• The proposed system improves the efficiency of analyzing complex EHR data for clinical research.
• It supports the exploration of patient pathways and simplifies the interpretation of deep learning model results.
• This approach can empower medical staff to test clinical hypotheses more effectively using big data.