Ethics and inclusion

This study constitutes human subject research. Institute Review Board (IRB) approval was obtained under Biomedical Research Alliance of New York (BRANY) protocol #21-08-508. As part of the BRANY IRB process, the protocol has been reviewed in accordance with the institutional guidelines. The BRANY waived the need for patient-informed consent and HIPAA authorization.

Data sources

This study is part of the National Institutes of Health (NIH) funded RECOVER Initiative (https://recovercovid.org/), which aims to learn about the long-term effects of COVID-19. The RECOVER pediatric cohort draws from EHRs provided by large national healthcare networks within the United States, covering regional catchment areas across 41 states. For this analysis, data were obtained from nineteen US children’s hospitals and health institutions, collectively covering over 30 million patients under 21, including more than 3 million individuals under 21 affected by COVID-19. The EHR data cover a wide range of healthcare interaction information routinely collected and stored by hospitals.

The institutions include Cincinnati Children’s Hospital Medical Center, Children’s Hospital of Philadelphia, Children’s Hospital of Colorado, University of Iowa Healthcare, Ann & Robert H. Lurie Children’s Hospital of Chicago, University of Michigan, University of Missouri, Montefiore, Medical University of South Carolina, Nationwide Children’s Hospital, Nemours Children’s Health System (in Delaware and Florida), OCHIN, Inc., Ochsner Health System, Ohio State University, Seattle Children’s Hospital, Stanford Children’s Health, Temple University, University of California, San Francisco, and Vanderbilt University Medical Center. The participating institutions represent a mix of public and private healthcare systems. While some of these institutions may serve insured individuals or fee-paying patients, many also provide care to uninsured or underinsured populations through government programs. Although some institutions may serve specific demographics or geographic areas, the overall dataset reflects a broad and heterogeneous sample of pediatric patients, representing various racial, ethnic, socioeconomic, and geographic backgrounds across the United States.

The EHR data was standardized to the PCORnet Common Data Model (CDM) and extracted from the RECOVER Database Version s10. More details are available in the Supplementary Materials Section S1.

Cohort construction and selection criteria

We conducted a retrospective study from March 1, 2020, to September 1, 2023, with a cohort entry period extending from March 1, 2020, to March 6, 2023, ensuring at least a 179-day follow-up for observing post-acute cardiovascular outcomes.

In our study, documented SARS-CoV-2 infections were defined by positive polymerase-chain-reaction (PCR), serology, antigen tests, or diagnoses of COVID-19, or diagnoses of PASC. The index date for SARS-CoV-2-positive patients was set as either the earliest date of positive tests, COVID-19 diagnoses, or 28 days before a PASC diagnosis. For SARS-CoV-2-negative patients, we required all tests to be negative, no evidence or diagnosis of COVID-19 or PASC during the study period, and at least one negative COVID-19 test within the cohort entry period. The index date for the comparator group was a randomly selected date from their negative tests to align the distribution of index dates between the two groups, controlling for time effects.

We included patients under 21, who had at least one healthcare visit within the baseline period, defined as 24 months to 7 days before the index date, and at least one encounter within the follow-up period, defined as 28 to 179 days after the index date, with certain institutions, to ensure active interaction with the healthcare system and adequate follow-up to assess post-acute outcomes.

Patients diagnosed with MIS-C, Kawasaki disease, CKD, or ESKD were excluded^{61,62,63,64,65,66}. Although MIS-C is considered a PASC, it was not included in this study’s post-acute cardiovascular outcomes for two reasons: children with MIS-C can be effectively treated with minimal post-acute cardiac sequelae⁶⁷, and MIS-C has been and continues to be extensively studied with significantly declining incidence over the past year⁶⁸.

Defining CHD

In this study, we included a range of CHD types, including: aortic valve stenosis (AVS), atrial septal defect (ASD), atrioventricular septal defect, bicuspid aortic valve, coarctation of aorta (CoA), dextro-transposition of the great arteries, double outlet right ventricle, Ebstein anomaly, hypoplastic left heart syndrome, interrupted aortic arch, mitral insufficiency, mitral stenosis, patent ductus arteriosus (PDA), pulmonary atresia, pulmonary valve stenosis, tetralogy of Fallot, total anomalous pulmonary venous connection (TAPVC), total anomalous pulmonary venous return (TAPVR), tricuspid atresia, truncus arteriosus, ventricular septal defect (VSD). Details of the code sets and the risk table detailing the breakdown of CHD types are available in Tables S1–S2.

Defining cardiovascular outcomes

We identified 18 post-acute cardiovascular outcomes for our study, including hypertension, atrial fibrillation, ventricular arrhythmias, atrial flutter, premature atrial or ventricular contractions, pericarditis, myocarditis, heart failure, cardiomyopathy, cardiac arrest, cardiogenic shock, pulmonary embolism, deep vein thrombosis, thrombophlebitis, thromboembolism, chest pain, palpitations, and syncope. These outcomes were assessed during the follow-up period in patients without a history of the specific condition during the baseline period.

We also grouped related cardiovascular outcomes into categories, including arrhythmias (atrial fibrillation, ventricular arrhythmias, atrial flutter, and premature atrial or ventricular contractions), inflammatory heart disease (pericarditis and myocarditis), other cardiac disorders (heart failure, cardiomyopathy, cardiac arrest, and cardiogenic shock), thrombotic disorders (pulmonary embolism, deep vein thrombosis, thrombophlebitis, and thromboembolism), cardiovascular-related symptoms (chest pain, palpitations, and syncope), and any cardiovascular outcome (any incident cardiovascular condition studied).

The cardiovascular outcomes analyzed in this study were identified using validated diagnostic codes (ICD-10-CM, ICD-10, ICD-9-CM, and SNOMED) applied across 19 institutions using the PCORnet Common Data Model (CDM)^69,70,71. The EHR data in this study were primarily entered by physicians and healthcare providers during routine clinical care. No free-text entries were used to identify cardiovascular outcomes in this study; all outcomes were derived directly from these validated diagnostic codes stored in the EHR systems. We used validated diagnostic codes confirmed by two board-certified pediatricians (DT, CF), with details of the code sets available in Supplementary Materials Table S1.

Covariates

We examined a detailed set of patient characteristics as measured confounders collected before cohort entry, to be adjusted through propensity score stratification²⁹ to balance the comparison groups. These included demographic factors, including age at index date, sex (female, male), and race/ethnicity (NHW, NHB, Hispanic, AAPI, Multiple, Other/Unknown); clinical factors, including obesity status, a chronic condition indicator defined by the Pediatric Medical Complexity Algorithm²⁸ (PMCA, no chronic condition, non-complex chronic condition, complex chronic condition), and a list of pre-existing chronic conditions^8,72; health care utilization factors collected 24 months to 7 days before index date, including the number of inpatient visits, outpatient visits, emergency department (ED) visits, unique medications, and negative COVID-19 tests (0, 1, 2, ≥3); vaccine information, including dosage of COVID-19 vaccine before index date (0, 1, ≥2) and interval since the last COVID-19 immunization (no vaccine, <4 months, ≥4 months); year-month of cohort entry (from March 2020 to March 2023); indicators from the 19 data-contributing sites.

Statistical analysis

We calculated the incidence of post-acute cardiovascular outcomes in SARS-CoV-2-positive and negative cohorts, stratified by CHD status. For each outcome, incidence rates were calculated by dividing new cases during the follow-up period by the total number of patients, excluding those with the specific outcome at baseline.

We presented distributions of preference scores—a transformation of propensity scores that accounts for prevalence differences between populations—to assess empirical equipoise^32,33. Preference scores, unlike traditional propensity scores, measure the relative likelihood of exposure versus non-exposure and adjust for differences in exposure prevalence. Empirical equipoise is achieved for SARS-CoV-2 positive and negative patients when the majority of individuals in both groups have preference scores ranging from 0.3 to 0.7, indicating substantial overlap in baseline characteristics^29,32. This overlap supports valid causal inference by ensuring comparisons are made within regions of the data with adequate representation.

To mitigate the effects of confounding, we used propensity score stratification^29,30,31 to adjust for a large number of measured confounders collected before the index date. We fitted a logistic regression model by regressing the response variable, SARS-CoV-2 infection status, on covariates, including demographic, clinical, and healthcare utilization factors as listed in the study variables. The predicted probabilities from the logistic regression model give the PS for each patient, representing their likelihood of belonging to the SARS-CoV-2 positive group given the observed covariates. The population was then stratified into five equally sized strata (quintiles) based on the distribution of propensity scores. Each stratum represented a distinct level of risk for SARS-CoV-2 positivity, making more comparable groups across strata. After stratification, we assessed the standardized mean difference (SMD) of each covariate between SARS-CoV-2 positive and negative patients, with an SMD of 0.1 or less indicating acceptable balance^29,73.

Within each stratum, the RR was estimated using a modified Poisson regression model for binary outcomes⁷⁴. To combine estimates across strata, a weighted average was applied, where the weights were proportional to the number of individuals in each stratum. This approach ensures that the combined RR reflects the population distribution accurately. RR is a collapsible measure, meaning the measure of association conditional on some factors remains consistent with the marginal measure collapsed over strata, which is crucial for accurate interpretation in clinical research^45,46.

We conducted an analysis stratified by immune status (immunocompromised versus not) within both the CHD and non-CHD groups was conducted. Additionally, subgroup analyses were performed across various demographic and clinical factors, including age (0–4, 5–11, 12–20 years), race/ethnicity (NHW, NHB, Hispanic), sex (male and female), obesity status (obese and non-obese), severity of acute COVID-19⁷⁵ (“non-severe” including asymptomatic and mild, “severe” including moderate and severe), and estimated time frames corresponding to dominant virus variants (pre-Delta, Delta, Omicron). Specifically, the pre-Delta variant spanned March 1, 2021, to June 30, 2021; Delta from July 1, 2021, to December 31, 2021; and Omicron from January 1, 2022, to March 6, 2023, with a minimum 179-day follow-up to observe PASC outcomes⁷⁶.

Sensitivity analysis

We conducted extensive sensitivity analyses to examine the robustness of our findings. We assessed death as an outcome in Section S4. While death is a major cardiac outcome, it is fortunately rare in the pediatric population. Negative control outcome experiments^29,47,48 were performed to calibrate the residual study bias from unmeasured confounders and systematic sources, in which the null hypothesis of no effect was believed to be true utilizing a list of 36 negative control outcomes determined by two board-certified pediatricians (DT, CF). The empirical null distribution and calibrated risks were reported in Supplementary Materials Section S5. Patients included solely based on a PASC diagnosis (Section S6) or those without any cardiovascular outcomes within the baseline period (Section S7) were excluded from additional analyses to examine the influence of potential selection bias and baseline differences. Given limited SARS-CoV-2 testing availability during the first wave of COVID-19 (March to May 2020), we conducted analyses excluding patients whose index dates fell within this period (Section S8). Furthermore, we excluded data from site L where the population is truncated at the ambulatory level (Section S9), to assess whether differences in data capture influenced the findings. To address the possibility that observed symptoms or manifestations attributed to PASC might overlap with those from other respiratory infections, we also compared the incidence of cardiovascular outcomes following SARS-CoV-2 infections with those following influenza or RSV infections (Section S10). All analyses were performed using R version 4.0.2. Statistical significance was set at a p-value < 0.05 (two-tailed).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.