Can longitudinal electronic health record data identify patients at higher risk of developing long COVID?

Abstract With hundreds of millions of COVID-19 infections to date, a considerable portion of the population has developed or will develop long COVID. Understanding the prevalence, risk factors, and healthcare costs of long COVID can be of significant societal importance. To investigate the utility of large-scale electronic health record (EHR) data in identifying and predicting long COVID, we analyzed a sample of 1.23 million participants from the National COVID Cohort Collaborative (N3C), a longitudinal EHR data repository from 80 sites in the US with over 8 million COVID-19 patients. We characterized the prevalence of long COVID using a few different types of definitions to illustrate their relative strengths and weaknesses. Then we developed machine learning models to predict the risk of developing long COVID using demographic factors and comorbidity in the EHR. The risk factors for long COVID include patient age; sex; smoking status; and comorbidities characterized by the Charlson Comorbidity Index (CCI). We were able to predict three types of long COVID with low to moderate levels of accuracy (AUC 0.599 – 0.734). We found that age and CCI were most predictive of long COVID diagnosis. Ongoing work includes applying the fair machine learning framework to the long COVID predictive models. We are implementing fairness and bias mitigation methods to model fitting through the following steps, selecting fairness metrics, preparing data and model, evaluating fairness metrics, applying bias mitigation methods to the dataset, and comparing model results and fairness metrics before and after the mitigation. The objective is to achieve equalized odds, a statistical notion that ensures classification algorithms do not discriminate against protected groups (such as sex and race/ethnicity). Results from the fairness-based machine learning will be included in the conference presentation.