Identification of post-COVID condition in a large population: a machine learning approach

Post-COVID condition (PCC) is a continuation or new development of symptoms long after recovery from acute illness of COVID-19. Administrative health data is a powerful source of information for large epidemiologic studies. However, PCC is undercoded in the health data; studies relying on PCC diagnosis can substantially underestimate disease prevalence. A machine learning (ML) model was developed to identify people with PCC from those with and without known COVID-19. The cross-sectional study included people with general practitioner visits between April 1, 2021, and March 31, 2022, in the province of Alberta, Canada. Predictors were derived from hospital admission, ambulatory care, and physician visit datasets; strategies were employed to minimize information loss and bias. Using diagnosed PCC as the reference standard, a penalized elastic-net logistic regression model was developed. Model development dataset included 3000 PCC cases and 27,437 non-cases. The model performed well in predicting the cases, with a receiver operating characteristic curve area of 0.96, 73% sensitivity, and 99% specificity. Applying the model to a population-based sample of 3.3 million identified 309,390 persons living with PCC, or a period prevalence estimate of 9.3%. Findings suggest an ML model approach can identify PCC from the health data with excellent accuracy. Our model was unique in incorporating individuals’ healthcare utilization information and trained with cases even without known COVID-19. Considering the underdetection of SARS-CoV-2 infections and undercoding of long-COVID in health data in many jurisdictions, the demonstrated approach would provide a practical alternative to identify persons living with PCC.