Abstract
BACKGROUND: The goal of this study was to evaluate whether International Classification of Diseases, 10th Revision (ICD-10), discharge data can be used to accurately differentiate primary coronavirus disease 2019 (COVID-19) hospitalizations, which are specifically due to COVID-19, from incidental COVID-19 hospitalizations for monitoring COVID-19 trends in a large county health department. We sought to explore the use of machine learning algorithms for enhancing surveillance capabilities in a local public health setting. METHODS: Discharge data for 5122 Santa Clara County hospitalizations with a positive severe acute respiratory syndrome coronavirus 2 polymerase chain reaction or antigen test occurring between December 15, 2021, and August 15, 2022, were used to train a series of models for classifying primary COVID-19 hospitalizations using chart review as a gold standard. Area under the receiver operating characteristic curve (AUROC) was used as the evaluation metric. RESULTS: Each model performed well when trained on the full set of available predictors. AUROC values ranged from 0.808 (random forest) to 0.818 (SuperLearner). After evaluating each model, we implemented a reporting process based on Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression, as the performance was comparable with SuperLearner and it had the advantage of being transparent and familiar to health department staff. CONCLUSIONS: In Santa Clara County, ICD-10 discharge data were successfully used to develop a low-burden method for monitoring primary COVID-19 hospitalization, demonstrating one way that predictive algorithms can help local health jurisdictions meet surveillance needs while minimizing manual effort.