Data source
We utilized the HealthJump stand-alone ambulatory database that compiles electronic health record (EHR) data including diagnoses, procedures, labs, vitals, medications, and histories sourced from participating members. This data is currently free and accessible with permissions from the COVID-19 Research Database consortium [19, 20]. This consortium, led by Datavant, brings together numerous private and public partners from various sectors to enable data access and promote knowledge sharing. The COVID-19 Research Database was established with approval from the Western Institutional Review Board and has an exemption from patient consent due to the use of deidentified data certified under HIPAA (Health Insurance Portability and Accountability Act of 1996), including both HIPAA limited data and non–HIPAA-covered data, and this exemption applies to all research conducted within the COVID-19 Research Database. In partnership with the COVID-19 Research Database, HealthJump enables authorized access to its database of records from various EMR platforms, and while all regions in the United States are represented, there are some limitations to this representation, particularly from New England and the West North Central states.
The process of obtaining access required the submission of a research protocol for review and approval by the consortium. After several procedural steps, we were granted access to the Amazon Web Services (AWS) cloud and the appropriate databases. Database software (Snowflake), statistical software (R, SAS, etc.) were made available within the AWS environment. The HealthJump data used for this study was last updated as of May 18th, 2023.
Study sample
The study included children from birth to 17 years old who had at least one confirmed diagnosis of COVID-19 (ICD-10 Code ‘U07.1x’) between March 2020 and May 2023 and with a date associated with that diagnosis code (Supplementary Figure 1).
Long COVID symptom and condition clusters
The main exposure was symptom and condition clusters defined by Rao et al. as being clinically and statistically significantly associated with Long COVID [13]. The symptoms that were identified were changes in smell or taste, hair loss, generalized pain, chest pain, abnormal liver enzymes, skin rashes, allergies, fatigue and malaise, fever and chills, cardio-respiratory symptoms, genitourinary symptoms, nausea and vomiting, and diarrhea. The conditions identified were myocarditis, acute respiratory distress syndrome, myositis, mental health conditions, disorders of the teeth and gingiva, ill-defined heart disease, fluid/electrolyte disturbances, thrombophlebitis and thromboembolism, acute kidney injury, tonsillitis, bronchiolitis, pneumonia, inflammatory skin conditions, obesity, motor disorders and gastroenteritis. Conditions and symptoms were classified using at least one ICD-10 diagnosis codes in the database between 30 and 180 days after initial COVID diagnosis. All other observed children were classified as not having Long COVID (Supplementary Fig. 1). We will refer to children having these symptoms or conditions as having “Long COVID” in the rest of the manuscript.
Other variables
We evaluated age, gender, race, region, and medical complexity as potential confounders. Medical complexity was defined using the complex chronic conditions (CCC) classification system developed by Feudtner et al. [21]. They defined CCC as “Any medical condition that can be reasonably expected to last at least 12 months (unless death intervenes) and to involve either several different organ systems or 1 organ system severely enough to require specialty pediatric care and probably some period of hospitalization in a tertiary care center,” and identified the ICD-10 diagnosis codes that were indicative of such complexity [21]. Analysis for pediatric CCC was completed in R using the pccc package as developed and described by Feinstein et al. [22].
Emergency department utilization visits or attendance
The outcome is ED use between 30 and 365 days after initial COVID diagnosis. ED visits were defined using the following CPT codes: 99,281, 99,282, 99,283, 99,284, 99,285.
Statistical analysis
Descriptive statistics were first employed to characterize the final study sample. We utilized t-tests, chi-squared tests, and Wilcoxon rank sum tests to compare ED visits among children indicating Long COVID to those without, based on variables such as gender, race, region, and medical complexity. We also evaluated the relative proportions of ED visits by computing percentages across population sizes by each Long COVID symptom.
To further explore the association between indication of Long COVID and ED use, we used a series of multivariable logistic regression models. Logistic regression is well-suited for modeling binary outcomes, making it especially useful for modeling ED utilization. Interpretability of such models is facilitated by the computation of odds ratios, allowing for a straightforward understanding of the impact of Long COVID symptoms on the likelihood of an ED visit. Models were built sequentially, starting with a crude model (Model 1), then adjusting for race and gender (Model 2), medical complexity (Model 3), year of COVID diagnosis (Model 4), and finally an interaction term with Long COVID and year of diagnosis (Model 5).
Among those with at least one symptom/condition, we assessed the association of specific symptoms and conditions with ED visits after adjusting for medical complexity. For model robustness, symptoms and conditions with low ED utilization were filtered out; we included symptoms and conditions affecting at least 30 patients and where the number of ED visits was at least 10. Adjusted odds ratios, 95% confidence intervals (95% CI), and p-values were calculated for each estimate. P-values < 0.05 were considered statistically significant. All analyses were conducted with SQL within the AWS environment and R/RStudio.
Add Comment