Identifying who has long COVID in the USA: a machine learning approach using N3C data

Area of Science:

• Health Informatics
• Machine Learning in Medicine
• Public Health

Background:

• Long COVID, or post-acute sequelae of SARS-CoV-2 infection, presents heterogeneous symptoms challenging precise definition.
• Electronic health records (EHRs) are vital for understanding Long COVID, as part of the NIH's RECOVER Initiative.
• Accurate identification of Long COVID patients is crucial for research and clinical management.

Purpose of the Study:

• Develop and validate machine learning models to identify potential Long COVID cases from EHR data.
• Enable efficient patient identification for clinical trials and specialized care.
• Provide a scalable method for Long COVID case ascertainment.

Main Methods:

• Utilized the National COVID Cohort Collaborative (N3C) EHR repository for a large adult patient cohort (n=1,793,604).
• Trained XGBoost models using demographics, healthcare utilization, diagnoses, and medications from 97,995 COVID-19 patients and 597 Long COVID clinic patients.
• Validated models on data from a fourth site, assessing performance across different patient subgroups (all, hospitalized, non-hospitalized).

Main Results:

• Achieved high accuracy in identifying potential Long COVID patients: AUC of 0.92 (all), 0.90 (hospitalized), and 0.85 (non-hospitalized).
• Key predictive features included healthcare utilization rates, patient age, dyspnea, and other EHR-documented diagnoses and medications.
• Shapley values were used to determine feature importance for model interpretability.

Conclusions:

• The developed models serve as a valuable proxy for identifying patients needing Long COVID specialty care.
• Facilitates the urgent need to identify Long COVID patients for clinical trials.
• Models are adaptable and can be retrained with new data sources to improve accuracy and meet specific study requirements.