Methods
We performed a retrospective cohort study using an all-payer administrative database comprised of patient encounter and billing data from 96 health systems (446 total hospitals) in the USA.
Patient and public involvement
Due to deidentification of the analytical sample, patients were not involved in study design or analysis.
Data source
Strata Decision Technology provides a financial planning, analytics and performance platform used by over 400 health systems (2000 total hospitals) in the USA. A subset of these health systems use StrataJazz Decision Support and participate in the StrataSphere analytics platform, providing deidentified patient encounter and billing data that are stored in an aggregated and anonymised database of financial information. These data include patient demographics, hospital admission and discharge timestamps, healthcare facility classification, discharge disposition, medical coding (including International Classification of Disease-10 (ICD-10) diagnosis codes and Current Procedural Terminology procedure codes), and itemised charges, and are updated daily. We used the charge data within this dataset as a proxy for healthcare costs20 and intensity of healthcare use.
Study population
The study included patients aged 18 years or older who were hospitalised with a primary diagnosis of either influenza or COVID-19, according to ICD-10 diagnoses codes (J09, J10, J11 for influenza and U07.1 for COVID-19). Because the onset of the COVID-19 pandemic was associated with a reciprocal decrease in influenza infections,21 22 we used asynchronous timeframes to generate cohorts for comparison (July 2018 to May 2021 for influenza, December 2019 to May 2021 for COVID-19), as seen in figure 1. Patients were excluded if any visit in the study period indicated a patient discharge code was ‘expired’ for any visit during the acute stage (n=26 134). Patients were also excluded if they were missing patient-level or hospital-level covariates (n = 2057), or if they were coinfected with influenza and COVID-19 according to ICD-10 diagnosis codes. Lastly, in order to directly compare patients with PASC to those with postviral complications from influenza, patients with no information available from the postacute stage (n=106 566) were excluded. Hospitalisation was defined by having room and board charges (based on uniform billing revenue codes between 0100 and 0179 or between 0190 and 0219) on the patient billing record.
Figure 1Timeline of study period. Admissions between July 2018 and May 2021 were included for influenza, and admissions between December 2019 and May 2021 were included for COVID-19. Billed charges for all encounters within the preacute period were summed and included as a covariate for linear modelling. Billed charges for encounters within the postperiod were the primary outcome.
Periods of comparison
In order to account for each patient’s baseline healthcare utilisation and charge generation, we collected data on charges before hospitalisation. Because acute infection-related charges may occur immediately preceding and following hospital admission, the 1-month periods before and after inpatient hospital admission and discharge were considered wash-out periods for healthcare expenditures related to the acute period. Therefore, the acute stage was defined as the period from 1 month (30 days) before the admission date for the patient’s first hospitalisation for COVID-19 or influenza to 1 month after the discharge date for that visit (figure 1). Charges during the acute stage were not included in the analysis. The preacute stage was defined as the 5-month period preceding the acute stage and the post-acute stage was defined as the 5-month period following the acute stage. We chose to study 6 months after COVID-19 diagnosis (counting acute and postacute periods) to include both national and international definitions of PASC5 23 and to be consistent with prior studies documenting outcomes after influenza and COVID-19.24 25 Only patients whose date of discharge was at least 5 months before the time of the data pull were included. Per-patient preacute charges were defined as the cumulative charges accrued from visits during the preacute stage. Per-patient postacute charges were defined as the cumulative charges accrued from visits during the postacute stage.
Covariate definitions
Mechanical ventilation was identified by the presence of ICD-10 procedure codes matching 5A1935Z, 5A1945Z or 5A1955Z. Primary and secondary ICD-10 diagnosis codes were used with the Elixhauser Comorbidity Software Refined for ICD-10-CM (V.2021.1) to identify comorbidities.26 For comorbidity measures requiring the present on admission (POA) indicator for assignment, POA codes matching Y (POA) or W (clinically undetermined) were included. In-hospital mortality was identified using the CMS discharge status codes matching 20 (expired), 41 (expired in a medical facility) or 42 (expired—place unknown).
The categorisation of hospitals as urban or rural was performed using geographic information system software, which provided precise determinations that do not rely on proxies such as counties or ZIP codes. Specifically, the physical address of each entity was geolocated against the most recently available Tiger/LINE Urban Areas shapefile from the Census Bureau. Entities lying within an ‘urbanised area’ (UA) according to the US Census Bureau’s definition of a contiguously built-up area of more than 50 000 people were deemed urban (otherwise rural). Other health system and hospital characteristics such as census region, bed size and operating expense were determined using data from a third-party vendor, Definitive Healthcare. The information from Definitive Healthcare was reconciled to each health system’s general ledger departments and entities to identify individual hospitals. In cases where crosswalks were indeterminate, additional resources such as the American Hospital Directory or contacts at the health system were referenced for verification.
Analysis
Demographic data, prevalence of medical comorbidities and hospitalisation characteristics were calculated for the COVID-19 and influenza cohort. We used linear regression models to examine the relationship between infection type (COVID-19 or influenza) and cumulative post-acute healthcare charges (postacute charges) among adults in the Strata-based cohort. Postacute charges were log-transformed to meet normality assumptions for the linear models. Analyses were stratified by age group (18–44, 45–64, 65+) and receipt of mechanical ventilation during hospitalisation due to hypothesised differences in age distribution and requirement for mechanical ventilation between the COVID-19 and influenza cohorts and anticipated confounding effects of these parameters on postacute charges. Selected age groups match US Census Bureau and Department of Health and Human Services criteria for midlife27 and older28 29 adults.
Each stratified model adjusted for patient-level covariates, including gender, race and ethnicity, medical comorbidities (derived from the Elixhauser comorbidities),14 preacute charges and date of admission. Preacute charges were modelled with a jump discontinuity at zero dollars to adjust for missing preacute charge information that could be attributable to either patients having no preacute care or patients seeking preacute and acute care in different hospital systems. Interaction terms were included to accommodate differential relationships between preacute and postacute charges by infection type. Elixhauser comorbidities were grouped by organ system as follows: presence/absence of cardiovascular disease (chronic heart failure, coagulopathy, peripheral vascular disease), lung disease (pulmonary circulatory disorders, chronic pulmonary disease), cancer (leukaemia, lymphoma, metastatic cancer, carcinoma in situ, solid malignancy), diabetes (controlled diabetes, uncontrolled diabetes), hypertension (controlled hypertension, uncontrolled hypertension), liver (mild liver disease, severe liver disease), renal (moderate renal failure, severe renal failure) and obesity.
All models also adjusted for characteristics of the patient’s acute hospital stay, including hospital census region, hospital bed size, urban hospital site, hospital system operating expense (over/under US$1 billion), admission date and length of stay (LOS) of patient’s first acute visit. Admission date was centred to January 2020 for analysis, and was modelled with a spline (knots at July 2019 and July 2020) to accommodate a potentially non-linear relationship between time and charges. Due to the markedly different temporal distribution of influenza and SARS-CoV-2 and the hypothesised differences in the populations susceptible to hospitalisation due to each disease, propensity score methods would not be expected to adequately balance the distribution of unobserved potential confounding variables in a way that would render inferences causal.
