A machine learning-based phenotype for long COVID in children: An EHR-based study from the RECOVER program

Area of Science:

• Pediatric Health Informatics
• Computational Epidemiology
• Machine Learning in Medicine

Background:

• Clinical understanding and definitions of pediatric Post-Acute Sequelae of SARS CoV-2 (PASC) are evolving.
• Reliable identification of PASC patients within health systems data is crucial for research and clinical care.
• Existing diagnostic codes may not fully capture the spectrum of PASC in children.

Purpose of the Study:

• To develop and validate a machine learning algorithm for identifying pediatric PASC patients in electronic health records (EHR).
• To distinguish between Multisystem Inflammatory Syndrome in Children (MIS-C) and non-MIS-C variants of PASC.
• To characterize PASC using diverse EHR features and understand contributing factors.

Main Methods:

• A cohort of pediatric patients with positive SARS-CoV-2 tests from the PEDSnet EHR network was analyzed.
• An XGBoost machine learning model was developed, utilizing conditions, procedures, diagnostic tests, and medications as features.
• Feature selection employed a tree-based scan statistic; model performance was validated using 5-fold cross-validation and SHAP values.

Main Results:

• A validated machine learning model was created to classify pediatric PASC patients.
• The model effectively utilizes EHR data, including diagnoses, medications, labs, and procedures, for PASC identification.
• SHAP values provided insights into feature importance for PASC classification.

Conclusions:

• The developed machine learning algorithm offers a reliable tool for identifying pediatric PASC in health systems data.
• The model supports precise patient identification for studies and efficient screening for clinical trials.
• Analysis of model features can enhance the understanding of PASC characteristics and associated factors in children.