Learning Clinical Concepts for Predicting Risk of Progression to Severe COVID-19

Area of Science:

• Computational epidemiology
• Biomedical informatics
• Clinical risk prediction

Background:

• Accurate identification of individuals at high risk for severe COVID-19 is critical for resource allocation and timely intervention.
• Electronic Health Record (EHR) data presents a rich source for predictive modeling but often suffers from under-coding, impacting model accuracy.
• Balancing model complexity (many features) with clinical interpretability (few features) is a key challenge in developing effective risk scores.

Purpose of the Study:

• To develop and validate high-performance survival models for predicting severe COVID-19 progression using EHR data.
• To investigate the utility of learning clinical concepts for improving risk prediction accuracy compared to using raw EHR features.
• To compare the performance of developed models against existing literature.

Main Methods:

• Development of two sets of risk scores: an unconstrained model using all available EHR features and a model utilizing a pipeline that learns clinical concepts before risk prediction.
• Utilizing survival analysis techniques to predict severe COVID-19 progression.
• External validation of models on out-of-sample data from subsequent time periods.

Main Results:

• The model incorporating learned clinical concepts achieved a higher C-index (0.858) compared to using corresponding raw features (0.844).
• The learned concepts approach demonstrated improved performance over the unconstrained model when evaluated on out-of-sample data.
• The developed models significantly outperformed previous works, with C-indices ranging from 0.844-0.872 compared to 0.598-0.810.

Conclusions:

• Learning clinical concepts from EHR data is an effective strategy to enhance the performance of severe COVID-19 risk prediction models.
• The developed risk scores offer a significant improvement over existing methods, providing a valuable tool for identifying high-risk individuals.
• This approach addresses the challenge of under-coded EHR data and the trade-off between model complexity and interpretability.