COVID-19 Mortality Prediction From Deep Learning in a Large Multistate Electronic Health Record and Laboratory...

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Computational Epidemiology

Background:

• COVID-19 presents diverse clinical outcomes, ranging from mild illness to fatal cardiopulmonary disease.
• Clinical data and laboratory tests are crucial for guiding patient treatment and stratification.
• Deep learning offers advanced methods for analyzing complex health datasets.

Purpose of the Study:

• To develop and validate a machine learning model for predicting COVID-19 patient mortality.
• To enable early identification of high-risk individuals shortly after SARS-CoV-2 infection confirmation.
• To leverage routine clinical and laboratory data for predictive modeling.

Main Methods:

• Construction of a large, geographically diverse dataset of COVID-19 patients from electronic health records.
• Evaluation of traditional machine learning models and recurrent neural network architectures (GRU-D).
• Training models using clinical covariates and laboratory measurements within 72 hours of positive COVID-19 test.

Main Results:

• The GRU-D model achieved a high area under the receiver operating characteristic (AUROC) curve of 0.938 during cross-validation.
• The model maintained strong predictive performance with reduced follow-up times (0.916 AUROC at 12 hours).
• Key predictors of mortality included age, comorbidity index, oxygen saturation, fibrinogen, and serum iron levels.

Conclusions:

• A deep learning approach (GRU-D) effectively predicts COVID-19 mortality risk.
• The model functions as an alert system using readily available clinical and lab data.
• This tool facilitates timely intervention for COVID-19 patients at high risk of death.