Trends in US Pediatric Hospital Admissions in 2020 Compared With the Decade Before the COVID-19 Pandemic

Key Points Question How have pediatric inpatient admission volumes changed in January to June 2020 compared with prior years? Findings This cross-sectional study of 5 424 688 admissions at 49 hospitals in the Pediatric Health Information Systems database used ensemble forecasting models to demonstrate differences between inpatient pediatric admissions in 2020 compared with prior years. There was a maximum 45.4% reduction in admissions in 2020, associated with a 27.7% reduction in hospital charges, with significant reductions in all examined diagnoses except for birth. Meaning In this study, inpatient pediatric admissions in 2020 were reduced across a heterogeneous range of diagnoses during the coronavirus disease 2019 pandemic.


Introduction
The first laboratory-confirmed case of coronavirus disease 2019 (COVID-19) was reported in the United States in January 2020. 1 In response, the United States declared a public health emergency on January 31, 2020, expanded airport screenings, required quarantines for travelers, and undertook public information efforts. 2,3 Case counts and fatalities increased slowly in January and February 2020 before achieving exponential growth in March 2020. 4 Individual states began additional mitigation strategies in the spring, including stay-at-home orders, which took effect from early March through June 2020. 5,6 These efforts were associated with decreased population movement, COVID-19 incidence, and mortality. 6,7 Additionally, the US Centers for Disease Control and Prevention issued recommendations for the use of face coverings on April 3, 2020, although enforcement varied by state. 3,8 Although COVID-19 has resulted in relatively few pediatric hospitalizations, 1,9,10 early reports indicate that it may have been associated with reduced pediatric admission rates in the United States 11 and worldwide. 12,13 A 2020 study from Brazil, 14 in which regional differences contributed to peak bronchiolitis between February and August, showed a 78% to 85% reduction in hospitalization in infants younger than 1 year for acute bronchiolitis in 2020 compared with previous years.
Many pediatric diagnoses warranting inpatient admission exhibit seasonal variation; prior work has characterized winter predominance in bronchiolitis, 15,16 pneumonia, 16,17 and Kawasaki syndrome 18 ; autumn predominance in asthma 16,19,20 ; and summer predominance in trauma. 21 These fluctuations in disease pattern are likely influenced by population susceptibility and behavior as well as various environmental factors. However, to our knowledge, the seasonal patterns in other pediatric admissions remain unexplored. The present study is a retrospective analysis of the Pediatric Health Information Systems (PHIS) database 22 to determine the seasonal patterns of a variety of common pediatric admission diagnoses and compare data from the previous decade with 2020 admissions. Given the changes in population behavior in 2020, 6,7 we hypothesized that pediatric admissions for a variety of diagnoses would be decreased in 2020 compared with prior years, except for admissions for birth. 23,24

Study Design and Participants
This was a retrospective cross-sectional study of patients from 51 children's hospitals across the United States participating in PHIS, an online, quality-controlled, anonymized, administrative database maintained by the Children's Hospital Association. 22,25 To ensure that changes in admission rates over time were reflective of changes in caseload, rather than expansion of the database by inclusion of new hospitals, the present study was limited to the 49 hospitals that have been providing data since 2010. Patients were eligible for inclusion if they were discharged between January 1, 2010, and June 30, 2020. There were no exclusion criteria. Patients were filtered according to their primary admission diagnoses based on International Classification of Diseases, Ninth Revision (ICD-9) or ICD-10 codes, depending on whether their admission was before or after October 1, 2015. Selected diagnoses were determined according to the percentage of total admissions in PHIS, disease pathophysiology, and seasonal distribution between 2010 and 2019. The full list of included ICD codes for each diagnosis group is included in eTable 1 in the Supplement. Outcomes included the number of monthly and annual admissions over time. This study was granted exemption by the

Statistical Analysis
Encounter-level clinical and administrative data were extracted from PHIS (admission and discharge dates, age, sex, race, ethnicity, diagnostic codes, insurance status, hospital length of stay, intensive care unit length of stay, complex chronic conditions, 27 mechanical ventilation, abstracted charges, and cost). Admissions were described with summary statistics. Hospital charges were adjusted by the Centers for Medicare & Medicaid Services wage/price index according to hospital zip code in PHIS. 22 Charge-over-time analyses were adjusted for the quarterly gross domestic product, provided by the Bureau of Economic Analysis and expressed in quarter 1, 2010, dollars. 28 Admission dates were grouped into month and year. Admission numbers were transformed into time series, databased on date of admission, and displayed graphically. Seasonality testing was performed with the Webel and Ollech method. 29 Seasonal trends in admission rates between 2010 and 2019 were displayed with locally estimated scatterplot smoothing. For visual comparison, the number of admissions in 2020 was displayed against the aforementioned models. For each diagnosis group, ensemble machine-learning forecasting models were created with autoregressive integrated moving average, 30 neural network, 31 and locally estimated scatterplot smoothing algorithms, 32 each weighted on time-series cross-validation. 33 Time-series data were arranged such that each model was trained on data from January 1, 2010, until June 30, 2019. Model forecasts were tested against actual admission numbers from July 1, 2019, until June 30, 2020. Model predictions were assessed with mean absolute percentage error, 33 and the error rates for July 2019 to December 2019 were compared against those for January 2020 to June 2020. To quantify whether social distancing had a statistically significant association with pediatric admission rates, the actual admission rate in 2019 vs 2020 was compared against the ensemble algorithm's 95% CI. 33 We conducted 2 exploratory analyses by repeating the aforementioned steps for patients with any diagnostic code for acute respiratory failure and for in-hospital mortality. All statistical analyses were performed with R Studio version 1.3.1073 and R versions 4.0.2 and 4.0.3 (R Project for Statistical Computing) with the following packages: Metrics, forecastHybrid, thief, forecast, cowplot, lubridate, forcats, stringr, dplyr, purrr, readr, tidyr, tibble, ggplot2, and tidyverse. 34 An α value of .05 was set as the threshold for statistical significance, and all tests were 2-tailed. The code used to create the manuscript is publicly available. 35

Demographic Characteristics
There were 49 hospitals with admission data in PHIS between January 1, 2010, and June 30, 2020.  There was a decrease in the number of admissions beginning in March 2020 compared with the

Seasonal Patterns in Selected Diagnoses
Admissions by diagnoses are shown in

Changes in Seasonal Patterns During the COVID-19 Pandemic
There were deviations from previous seasonal patterns in 2020, as shown in Figure 2 (denoted in red). Admissions for bronchiolitis and S pneumoniae reached previous summer rates by April 2020.
Admissions for trauma did not exhibit a typical increase in spring 2020. Admissions for appendicitis, atrial septal defects, asthma, coarctation of the aorta, dehydration, hypoplastic left heart syndrome, Kawasaki syndrome, mental health conditions, sepsis, and tetralogy of Fallot decreased below expected levels. Birth rates were unaffected. Changes in seasonal patterns identified with ensemble time-series forecasting models are shown in Figure 3 and eFigure 7 in the Supplement. Overall, models accurately predicted admission rates from July until December 2019 but not from January to June 2020 (mean absolute percentage error range, 3% to 21.2% vs 8.5% to 1111.7%, respectively; within the 95% CI of the models. The admission rates for atrial septal defect and hypoplastic left heart syndrome were below the 95% CI predicted by the models in August and December 2019, respectively. All other conditions remained within the model 95% CIs between July and December 2019. Admissions for birth remained within the model 95% CI between January and June 2020.
Every other condition decreased below the model 95% CI between January and June 2020.

Exploratory Analyses
Diagnoses including any cause of acute respiratory failure displayed significant winter predominance.
In-hospital mortality did not meet criteria for

Discussion
To our knowledge, the present study is the first to use a large database to model pediatric admission volumes during the COVID-19 pandemic for various diagnoses compared with prior years. Using this approach, we identified seasonal patterns in a diverse group of pediatric conditions. Before 2020, pediatric admissions overall displayed winter predominance, associated with an increase in infectious respiratory conditions, such as bronchiolitis and pneumonia. Summer-predominant conditions included trauma and admissions for semielective surgical conditions (eg, atrial septal defects, hypoplastic left heart syndrome in children aged >1 year). The present study is in agreement with several prior studies documenting seasonality in bronchiolitis, 15,16 pneumonia, 16,17 asthma, 16,19,20 Kawasaki syndrome, 18 and trauma. 21 However, we additionally noted novel seasonality for diabetic ketoacidosis and mental health admissions and the absence of seasonality in admissions for cardiac arrest and coarctation of the aorta. The seasonality of diabetic ketoacidosis was associated with children aged 5 years or older and may be reflective of the efforts of the American Diabetes Association to incorporate school personnel in the diabetes medical management plans for these patients. 36 Although COVID-19 has resulted in few pediatric admissions and mortalities compared with those for adults, 1,9,10  our findings, with more than 95% reduction in respiratory syncytial virus and influenza detection in Western Australia through winter 2020. 40 Taken together, these studies suggest that rates of infectious pediatric respiratory conditions are potentially modifiable by changes in human behavior.
Similarly, pediatric trauma admissions are often sports related, 41 and the COVID-19 pandemic has been shown to be associated with decreased childhood physical activity; thus, reductions in trauma admissions were anticipated. 42 However, reductions in mental health admissions are surprising, given limited literature suggesting that pandemic-related stressors may worsen mental health in both the general population and those with preexisting psychiatric conditions. 43,44 This may represent an opportunity for changes in the delivery of mental health care. 45 The present study cannot delineate which changes in admission rates represent decreases in incidence, hospital avoidance, or unmet care needs. One inpatient survey suggested that greater than one-third of parents delayed seeking medical care for their child because of fears surrounding COVID-19, 46 and case series have reported delayed presentations leading to harm. 47,48 Similarly, although admissions for repair of atrial septal defects or later-stage repairs of hypoplastic left heart syndrome may safely be deferred for a short period, it is unclear whether reductions in admission rates represent purely elective surgical delays or the possibility of decreased recognition of congenital heart disease because of decreased contact with health care; small case series of increases in sudden death in adults with congenital heart disease without contact with health care institutions during the COVID-19 pandemic have been reported. 49 Taken together, these findings are worrisome, although not definitive, in that unmet health care needs may be accumulating in the pediatric population as a result of decreased health care interactions.
Reduced patient volumes during early 2020 have reportedly contributed to financial instability for many hospitals and health systems. 50 The reduction in charges observed in the present study Characterizing alterations in usual patient admission patterns aids in anticipating the potential ramifications of recent national policies and can also help in crafting responses in a variety of areas.
The present study highlights the strengths of an ensemble machine-learning approach to model complex time-series data displaying a variety of seasonal relationships. As complexity in medical information has increased, there has been progressive interest in the application of machine learning to facilitate clinical decision-making. Such algorithms have been studied in a diverse array of applications, including sepsis prediction, 52 COVID-19 prognosis, 53 population health, 54 and cyberbullying. 55 In the present study, before the COVID-19 pandemic, 6,7 ensemble time-series forecasting models accurately predicted 70 of 72 monthly pediatric admission rates (97.2%) between July 2019 and December 2019. Quantification of the seasonality of various admission types in pediatrics has applicability for hospital, intensive care unit, and ventilator capacity planning as well as trainee education. According to the present study, pediatricians and trainees assigned to inpatient wards in the winter will experience higher-than-average case burden and gain greater experience in managing pneumonia and bronchiolitis, whereas those assigned to emergency departments in the summer will achieve greater experience in the management of pediatric trauma. Such knowledge can be harnessed in the development of training curricula that support experience across the spectrum of pediatric illness.

JAMA Network Open | Pediatrics
Trends in US Pediatric Hospital Admissions in 2020 vs Before the Pandemic

Limitations
This study has important limitations. As with any retrospective database analysis, admission diagnoses are vulnerable to misclassification owing to coding error. Although including greater than 5 000 000 inpatient encounters, this study is limited to 49 hospitals that consistently provided data to PHIS between January 2010 and June 2020 and may not be representative of national pediatric admission trends. Because of data availability, data from some admissions from late June 2020 are not yet available. This may have resulted in undercounting of admissions from June. However, hospitalizations for birth did not decrease below the 95% CIs for the forecasting model. It is also possible that some hospitals have been delayed in reporting admissions to PHIS because of the pandemic, although this is not a known data quality issue. Owing to data availability, only admissions between January and June 2020 were analyzed regarding the effects of social distancing. Further study is warranted as additional data become available.

Conclusions
In this study, common pediatric admission diagnoses exhibited seasonal variation. These trends were effectively modeled with ensemble time-series forecasting algorithms. The overall number of pediatric admissions in PHIS decreased between January 2020 and June 2020. This trend held through a wide variety of diagnoses. The reduction in pediatric admissions may be representative of unmet needs in pediatric care during the COVID-19 pandemic.