Predicting Long COVID in the National COVID Cohort Collaborative Using Super Learner: Cohort Study

Area of Science:

• Computational biology
• Epidemiology
• Health informatics

Background:

• Postacute sequelae of COVID-19 (PASC), or long COVID, presents diverse long-term symptoms across biological systems.
• Identifying PASC risk factors and etiology is challenging due to symptom heterogeneity.
• Predictive characteristics for PASC are valuable for early identification and prevention.

Purpose of the Study:

• To predict individual risk of PASC diagnosis using clinical data.
• To identify key predictive factors for PASC development.
• To leverage electronic health records for PASC risk assessment.

Main Methods:

• Utilized Super Learner, an ensemble machine learning algorithm, combining gradient boosting and random forest.
• Analyzed data from 55,257 patients (1:4 PASC to control ratio) from the National COVID Cohort Collaborative.
• Evaluated variable importance using Shapley values across individual features, temporal windows, and clinical domains.

Main Results:

• Achieved accurate PASC diagnosis prediction with an area under the curve of 0.874.
• Top predictors included observation period length, healthcare interactions during acute COVID-19, and lower respiratory infection.
• Baseline characteristics were most predictive, followed by health care use, demographics, and respiratory factors.

Conclusions:

• Developed an open-source method for PASC risk prediction using electronic health record data.
• Health care use emerged as a strong predictor, necessitating careful consideration in observational studies.
• Early risk assessment before acute COVID-19, focusing on baseline and respiratory factors, can enhance interventions.