Deep EHR Spotlight: a Framework and Mechanism to Highlight Events in Electronic Health Records for Explainable...

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Healthcare
• Data Visualization

Background:

• Electronic Health Records (EHRs) generate vast clinical data, promising research advancements.
• Deep learning models excel at EHR predictive analytics but often lack transparency and require extensive preprocessing.
• Heterogeneous EHR data (text, numbers, time series) complicates visualization and interpretability.

Purpose of the Study:

• To propose a deep learning framework for encoding patient pathways from EHRs into visual representations.
• To improve the transparency and intelligibility of predictive models using EHR data.
• To enable complex, sequential outcome predictions with enhanced visualization.

Main Methods:

• Encoding patient clinical pathways from EHR data into image formats.
• Utilizing a deep attention mechanism for highlighting significant events within pathway images.
• Developing a framework for multi-sequential outcome prediction with visual interpretability.

Main Results:

• The proposed framework successfully encodes patient pathways into images.
• The deep attention mechanism effectively visualizes important clinical events.
• The model demonstrates capability for predicting multiple sequential outcomes with enhanced intelligibility.

Conclusions:

• The developed deep learning framework offers a novel approach to visualize and interpret EHR data.
• This method addresses the transparency and interpretability challenges in deep learning for clinical informatics.
• The framework facilitates more insightful predictions from complex, multi-modal EHR data.