A Multi-Layered GRU Model for COVID-19 Patient Representation and Phenotyping from Large-Scale EHR Data

Area of Science:

• Computational biology
• Health informatics
• Epidemiology

Background:

• The COVID-19 pandemic caused severe healthcare resource shortages.
• Predicting patient outcomes and phenotypes is crucial for efficient resource allocation and targeted treatment.

Purpose of the Study:

• To develop and evaluate a data-driven model for predicting COVID-19 patient outcomes.
• To identify distinct patient phenotypes using electronic health record data.
• To provide computational tools for proactive healthcare resource allocation.

Main Methods:

• Developed a multi-layered gated recurrent units-based model (mGRU-CP).
• Leveraged electronic health record (EHR) data from the COVID-19 Research Data Commons.
• Compared mGRU-CP against four state-of-the-art baseline methods using three patient feature sets.

Main Results:

• mGRU-CP achieved competitive or superior performance compared to baseline methods across all settings.
• Learned patient embeddings from mGRU-CP enabled meaningful patient phenotyping.
• The model effectively predicted patient survival probabilities.

Conclusions:

• The mGRU-CP model provides a robust computational tool for predicting COVID-19 patient outcomes.
• Patient phenotyping using learned embeddings enhances understanding of mortality.
• This approach can inform healthcare resource allocation strategies for future pandemics.