Who has long-COVID? A big data approach [preprint]

Background Post-acute sequelae of SARS-CoV-2 infection (PASC), otherwise known as long-COVID, have severely impacted recovery from the pandemic for patients and society alike. This new disease is characterized by evolving, heterogeneous symptoms, making it challenging to derive an unambiguous long-COVID definition. Electronic health record (EHR) studies are a critical element of the NIH Researching COVID to Enhance Recovery (RECOVER) Initiative, which is addressing the urgent need to understand PASC, accurately identify who has PASC, and identify treatments. Methods Using the National COVID Cohort Collaborative’s (N3C) EHR repository, we developed XGBoost machine learning (ML) models to identify potential long-COVID patients. We examined demographics, healthcare utilization, diagnoses, and medications for 97,995 adult COVID-19 patients. We used these features and 597 long-COVID clinic patients to train three ML models to identify potential long-COVID patients among (1) all COVID-19 patients, (2) patients hospitalized with COVID-19, and (3) patients who had COVID-19 but were not hospitalized. Findings Our models identified potential long-COVID patients with high accuracy, achieving areas under the receiver operator characteristic curve of 0.91 (all patients), 0.90 (hospitalized); and 0.85 (non-hospitalized). Important features include rate of healthcare utilization, patient age, dyspnea, and other diagnosis and medication information available within the EHR. Applying the “all patients” model to the larger N3C cohort identified 100,263 potential long-COVID patients. Interpretation Patients flagged by our models can be interpreted as “patients likely to be referred to or seek care at a long-COVID specialty clinic,” an essential proxy for long-COVID diagnosis in the current absence of a definition. We also achieve the urgent goal of identifying potential long-COVID patients for clinical trials. As more data sources are identified, the models can be retrained and tuned based on study needs. Funding This study was funded by NCATS and NIH through the RECOVER Initiative