Evaluation of machine learning for predicting COVID-19 outcomes from a national electronic medical records database

Abstract

Objective: When novel diseases such as COVID-19 emerge, predictors of clinical outcomes might be unknown. Using data from electronic medical records (EMR) allows evaluation of potential predictors without selecting specific features a priori for a model. We evaluated different machine learning models for predicting outcomes among COVID-19 inpatients using raw EMR data. Materials and Methods: In Premier Healthcare Data Special Release: COVID-19 Edition (PHD-SR COVID-19, release date March, 24 2021), we included patients admitted with COVID-19 during February 2020 through April 2021 and built time-ordered medical histories. Setting the prediction horizon at 24 hours into the first COVID-19 inpatient visit, we aimed to predict intensive care unit (ICU) admission , hyperinflammatory syndrome (HS), and death. We evaluated the following models: L2-penalized logistic regression, random forest, gradient boosting classifier, deep averaging network, and recurrent neural network with a long short-term memory cell. Results: There were 57,355 COVID-19 patients identified in PHD-SR COVID-19. ICU admission was the easiest outcome to predict (best AUC=79%), and HS was the hardest to predict (best AUC=70%). Models performed similarly within each outcome. Discussion: Although the models learned to attend to meaningful clinical information, they performed similarly, suggesting performance limitations are inherent to the data. Conclusion: Predictive models using raw EMR data are promising because they can use many observations and encompass a large feature space; however, traditional and deep learning models may perform similarly when few features are available at the individual patient level.