Clinical decision support using pseudo-notes from multiple streams of EHR data.

Area of Science:

• Artificial Intelligence
• Clinical Informatics
• Machine Learning

Background:

• Electronic Health Records (EHR) integrate diverse data, posing challenges for analysis.
• Heterogeneous EHR data requires sophisticated methods for effective clinical decision support.

Purpose of the Study:

• To introduce the Multiple Embedding Model for EHR (MEME), a novel deep learning framework.
• To enhance clinical decision support by leveraging heterogeneous EHR data.

Main Methods:

• MEME converts tabular EHR data into "pseudo-notes" for compatibility with language foundation models.
• The framework embeds EHR domains separately and uses self-attention to learn contextual importance.
• A study analyzed 400,019 emergency department visits.

Main Results:

• MEME accurately predicted emergency department disposition, discharge location, intensive care needs, and mortality.
• The model surpassed traditional machine learning, EHR foundation models, and GPT-4 prompting strategies.
• MEME demonstrated strong few-shot learning on an external, unstandardized EHR database.

Conclusions:

• MEME offers a robust deep learning approach for clinical decision support using heterogeneous EHR data.
• The framework's pseudo-note conversion and multi-embedding strategy improve predictive accuracy and adaptability.
• MEME shows promise for advancing AI in healthcare, particularly in handling diverse and unstandardized data.