Estimating the Burden of Influenza Hospitalizations Across Multiple Seasons Using Capture-Recapture

Abstract

Background

Influenza remains an important cause of hospitalizations in the United States. Estimating the number of influenza hospitalizations is vital for public health decision making. Combining existing surveillance systems through capture-recapture methods allows for more comprehensive burden estimations.

Methods

Data from independent surveillance systems were combined using capture-recapture methods to estimate influenza hospitalization rates for children and adults in Middle Tennessee during consecutive influenza seasons from 2016–2017 through 2019–2020. The Emerging Infections Program (EIP) identified cases through surveillance of laboratory results for hospitalized children and adults. The Hospitalized Adult Influenza Vaccine Effectiveness Network (HAIVEN) and New Vaccine Surveillance Network (NVSN) recruited hospitalized patients with respiratory symptoms or fever. Population-based influenza rates and the proportion of cases detected by each surveillance system were calculated.

Results

Estimated overall influenza hospitalization rates ranged from 23 influenza-related hospitalizations per 10 000 persons in 2016–2017 to 40 per 10 000 persons in 2017–2018. Adults aged ≥65 years had the highest hospitalization rates across seasons and experienced a rate of 170 hospitalizations per 10 000 persons during the 2017–2018 season. EIP consistently identified a higher proportion of influenza cases for adults and children compared with HAIVEN and NVSN, respectively.

Conclusions

Current surveillance systems underestimate the influenza burden. Capture-recapture provides an alternative approach to use data from independent surveillance systems and complement population-based burden estimates.

Capture-recapture provides an alternative approach to examine the detection of cases by independent influenza surveillance systems. Our capture-recapture estimates were consistently higher across all seasons than what each surveillance system independently identified, indicating that current surveillance underestimates cases.

Every year, influenza viruses cause an estimated 9.2–35.6 million influenza-related infections, 139 000–708 000 hospitalizations, and 12 000–56 000 deaths in the United States [1]. Influenza surveillance systems measure the seasonal burden of influenza, identify the predominant circulating strains, describe the populations most at risk from illness, and enable evaluations of vaccine effectiveness each season. Diagnosing influenza illness based on clinical symptoms alone is inaccurate and unreliable, and laboratory testing is required for confirmation of influenza virus infection. Therefore, surveillance systems that rely on clinician-driven testing for influenza may underestimate the burden of influenza.

While surveillance systems aim to monitor influenza activity and burden, it is logistically challenging to obtain complete burden estimates with direct enumeration of each infected case as many cases may not seek medical care and even among those who seek medical care, not all will be tested, resulting in missed cases. For distinct surveillance systems that operate in the same geographical area, alternative methods could be used to integrate their individual detections to obtain a more comprehensive estimate of the disease burden. Capture-recapture methods can improve estimation of disease burden by using 2 or more independent surveillance systems that operate concurrently in a population [2–4]. We applied capture-recapture methods using data from 3 hospital-based surveillance systems operating in Middle Tennessee to calculate influenza hospitalization rates for both children and adults during consecutive influenza seasons from 2016–2017 through 2019–2020.

METHODS

Influenza Hospital-Based Surveillance Systems

The current capture-recapture analyses used data from 3 surveillance systems that identified patients admitted to the hospital with an influenza illness and operated concurrently in Middle Tennessee during consecutive influenza seasons (2016–2020). The Emerging Infections Program (EIP) and the Hospitalized Adult Influenza Vaccine Effectiveness Network (HAIVEN) provided data for adults, whereas EIP and the New Vaccine Surveillance Network (NVSN) provided data for children. Individual patients were identified in each surveillance system and overlapping cases (ie, detected by both systems) were identified based on first and last name, date of birth, county of residence, admission hospital, and date of admission. The Institutional Review Board (IRB) for the Vanderbilt University Medical Center (VUMC) and the Tennessee Department of Public Health approved these analyses. EIP was approved by the VUMC and the State of Tennessee IRB, HAIVEN was approved by the VUMC and Sterling IRB, and NVSN was approved by the VUMC IRB.

The EIP conducts active laboratory and population-based infectious disease surveillance through a collaboration between the Centers for Disease Control and Prevention (CDC), state health departments, and academic institutions. EIP identifies all adult and pediatric patients hospitalized with a positive influenza laboratory test within 14 days prior to or during hospitalization, by either rapid diagnostic influenza test, reverse-transcription polymerase chain reaction (RT-PCR), or direct fluorescent antibody who reside within a defined surveillance area. Influenza testing is determined by treating providers. In Tennessee, surveillance is conducted at 18 hospitals encompassing a catchment area of 8 counties comprising 1.7 million residents.

HAIVEN identifies adult patients who are hospitalized with symptoms of acute respiratory illness, specifically cough, at 3 hospitals in the geographic catchment area through electronic medical record screening review. After informed consent was obtained, nasal and oropharyngeal swabs were collected by research staff, regardless of clinician-driven testing. Specimens were tested by RT-PCR for influenza in a research laboratory. HAIVEN had 3 participating hospitals where enrollment occurred approximately 5 days per week but attempted to include daily admission for 7 days per week. Patient participation was not limited by a specific catchment area, but home location was ascertained for further analyses.

NVSN performs active surveillance to identify children (<18 years of age) hospitalized with acute respiratory illness in 1 pediatric hospital in the catchment area. Children with fever and/or respiratory symptoms at 1 children's hospital were identified by screening electronic medical records and approached for enrollment within 48 hours of admission. After informed consent was obtained, nasal and oropharyngeal swabs were collected and tested for respiratory viruses by RT-PCR, including influenza, in a research laboratory [5]. Similar to HAIVEN, specimens were collected from enrolled patients, regardless of clinician-driven testing. Additionally, HAIVEN and NVSN samples were analyzed in the same laboratory. Surveillance was conducted daily throughout the influenza season [3]. Patient participation was not limited by a specific catchment area, but home location was ascertained for further analyses.

Study Population

The study catchment area included Davidson County and the surrounding 7 counties (Cheatham, Dickson, Robertson, Rutherford, Sumner, Williamson, and Wilson) where EIP operated. HAIVEN and NSVN participants were limited to those who resided in the above-defined catchment area. As HAIVEN and NVSN did not perform surveillance at all EIP hospitals, the hospitals were limited to those participating in HAIVEN and NVSN. We included all influenza cases and did not stratify by subtype but for completeness, we describe the predominant circulating strain for each season in the footnotes of Table 1.

Table 1.

Estimated Influenza Hospitalization Rates per 10 000 Persons by Age Group in Middle Tennessee Across 4 Influenza Seasons Based on Capture-Recapture Estimates

Age Group	2016–2017a		2017–2018a		2018–2019a		2019–2020a
	Rate per 10 000	(95% CI)	Rate per 10 000	(95% CI)	Rate per 10 000	(95% CI)	Rate per 10 000	(95% CI)
0–17 y	4	(3–9)	6	(4–15)	22	(14–42)	8	(7–11)
18–49 y	9	(7–11)	14	(12–17)	6	(5–8)	11	(10–13)
50–64 y	34	(29–43)	57	(49–70)	32	(28–39)	36	(33–44)
≥65 y	93	(79–118)	170	(147–208)	108	(94–133)	106	(95–128)
Total	23	(19–30)	40	(34–51)	28	(23–38)	28	(25–34)

Abbreviation: CI, confidence interval.

^aDominant circulating virus was A(H3N2) in 2016–2017 [9], A(H3N2) in 2017–2018 [10], A(H1N1) in 2018–2019 [11], and A(H1N1) in 2019–2020 [12].

Four influenza seasons were included in these analyses: 2016–2017, 2017–2018, 2018–2019, and 2019–2020. Surveillance systems operated during each influenza season from 1 October through 30 April. However, during the 2019–2020 season, the last few weeks of surveillance activities were likely impacted by implementation of nonpharmaceutical interventions for the coronavirus disease 2019 (COVID-19) pandemic.

Statistical Analyses

The total number of people hospitalized with influenza was estimated using the Chapman capture-recapture method as a less biased estimate when operating with small numbers than other capture-recapture methods (Figure 1) [6]. The Chapman method estimates the number of people hospitalized with influenza N within each season, by calculating $\hat{N}$. The calculation utilized the number of people hospitalized with influenza detected by 2 independent surveillance systems. Estimates were computed separately for children and adults. The first surveillance system, either HAIVEN or NVSN (for either adults or children), captured n1 influenza hospitalizations. The second surveillance system, EIP, captured n2 influenza hospitalizations for the corresponding age group. The number of influenza hospitalizations m2, were those captured by both surveillance systems. The probability of capture by one of these surveillance systems was assumed to be independent of the probability of capture by the other surveillance system as both HAIVEN and NVSN actively enrolled and tested participants hospitalized with respiratory symptoms, regardless of the treating clinician testing decision, whereas EIP is a surveillance system that relies on results from clinician-driven testing. We used transformed logit confidence intervals (CIs) to ensure proper coverage for capture-recapture estimates based on small counts [7]. To examine the yield of each surveillance system, we calculated the proportion of influenza hospitalizations detected by each surveillance system as the number of influenza hospitalizations detected by each surveillance system divided by the total number of hospitalizations estimated by the capture-recapture method.

Figure 1.

Capture-recapture estimation using data from 2 independent surveillance sources using the Chapman method. Abbreviations: EIP, Emerging Infections Program; HAIVEN, Hospitalized Adult Influenza Vaccine Effectiveness Network; NVSN, New Vaccine Surveillance Network.

After estimating the total number of people hospitalized with influenza through the capture-recapture analyses, we used the estimated numbers to calculate hospitalization rates overall and for each influenza season. For each season, hospitalization rates were calculated using the capture-recapture results, hospital market share data, and census population estimates. Market share was used to determine the proportion of hospitalizations for acute respiratory illnesses in the county that occurred in the surveillance hospitals. Hospitalization data came from the Hospital Discharge Data System, which is a registry of all hospitalizations in the State of Tennessee and includes demographic information including addresses, hospital information, and hospital discharge diagnosis codes. To estimate the total number of influenza hospitalizations in the study area comprised of the 8 counties, we weighted the capture-recapture estimates to account for the corresponding market share estimates, and divided these estimated totals by the corresponding census population estimates as denominators for calculation of population-based hospitalization rates [8]. Estimates were computed separately for children and for adults and stratified by season. Age-specific estimates were reported along with 95% CIs for children aged 0–17 years and adults aged 18–49, 50–64, and ≥65 years. Analyses were performed using Stata version 16.1 (StataCorp, College Station, Texas) and SAS Enterprise Guide 8.1 (SAS Institute, Cary, North Carolina) software.

RESULTS

Over the 4 surveillance seasons, EIP identified 910 hospitalized adults and 240 hospitalized children at hospitals participating in both HAIVEN and NVSN, respectively. HAIVEN identified 345 hospitalized adults and NSVN identified 235 hospitalized children (Table 2).

Table 2.

Estimated Capture-Recapture Hospitalizations Detected by Concurrent Surveillance Systems Across 4 Influenza Seasons, Middle Tennessee

Age Category	Surveillance System	2016–2017	2017–2018	2018–2019	2019–2020
Adult	HAIVEN	76	102	87	80
	EIP	167	315	191	237
	Overlapped	38	52	47	50
	Capture-recapture	331	613	351	377
Pediatric	NVSN	20	34	135	46
	EIP	41	31	46	122
	Overlapped	6	5	9	24
	Capture-recapture	125	186	638	230

Abbreviations: EIP, Emerging Infections Program; HAIVEN, Hospitalized Adult Influenza Vaccine Effectiveness Network; NVSN, New Vaccine Surveillance Network.

The capture-recapture analyses were stratified by season and between pediatric and adult hospitalizations (Figure 2). For example, the capture-recapture analysis estimated that the EIP and NVSN surveillance systems missed 70 influenza-related pediatric hospitalizations during the 2016–2017 influenza season, bringing the total number of estimated hospitalizations at the children's hospital for the season up to 125 (95% CI, 83–271) (Supplementary Table 1). Pediatric capture-recapture estimates were highest for the 2018–2019 season with 638 hospitalizations (95% CI, 410–1216), followed by the 2019–2020 season with 230 hospitalizations (95% CI, 191–315). Adult capture-recapture estimates were highest in the 2017–2018 season with 613 hospitalizations (95% CI, 531–751) between the 3 hospitals, followed by the 2019–2020 season with 377 hospitalizations (95% CI, 335–454) (Supplementary Table 2).

Figure 2.

Influenza hospitalizations by capture-recapture estimates and surveillance systems for pediatric (A) and adult (B) cases in Middle Tennessee across 4 influenza seasons. Abbreviations: EIP, Emerging Infections Program; HAIVEN, Hospitalized Adult Influenza Vaccine Effectiveness Network; NVSN, New Vaccine Surveillance Network. Transformed logit confidence intervals were used for capture-recapture; Poisson confidence intervals were used for EIP and HAIVEN/NVSN.

Using capture-recapture estimates as a reference, the proportion of influenza hospitalizations detected by each surveillance system was examined. EIP consistently detected more adult influenza hospitalizations than HAIVEN (Table 3). The proportion of cases detected by the EIP ranged from 50.4% in the 2016–2017 season to 62.9% in the 2019–20 season. In contrast, the proportion of cases detected by HAIVEN ranged from 16.6% in the 2017–2018 season to 24.8% in the 2018–2019 season. Among children, EIP detected more hospitalizations in the 2016–2017 and 2019–2020 seasons, but NVSN detected more cases in the 2017–2018 and 2018–2019 seasons. The proportion of cases detected by EIP ranged from 7.2% in the 2018–2019 season to 53.0% during the 2019–2020 season. The proportion of pediatric cases detected by NVSN ranged from 16.0% during the 2016–2017 season to 21.2% in the 2018–2019 season.

Table 3.

Proportion^a of Estimated Capture-Recapture Hospitalizations Detected by Surveillance System and Influenza Season in Middle Tennessee

Age Category	Surveillance System	2016–2017		2017–2018		2018–2019		2019–2020
		Proportion	(95% CI)	Proportion	(95% CI)	Proportion	(95% CI)	Proportion	(95% CI)
Adult	HAIVEN	23.0%	(18.4–27.5)	16.6%	(13.7–19.6)	24.8%	(20.3–29.3)	21.2%	(17.1–25.3)
	EIP	50.4%	(45.1–55.8)	51.4%	(47.4–55.3)	54.4%	(49.2–59.6)	62.9%	(58.0–67.7)
Pediatric	NVSN	16.0%	(9.6–22.4)	18.3%	(12.7–23.8)	21.2%	(18.0–24.3)	20.0%	(14.8–25.2)
	EIP	32.8%	(24.6–41.0)	16.7%	(11.3–22.0)	7.2%	(5.2–9.2)	53.0%	(46.6–59.5)

Abbreviations: CI, confidence interval; EIP, Emerging Infections Program; HAIVEN, Hospitalized Adult Influenza Vaccine Effectiveness Network; NVSN, New Vaccine Surveillance Network.

^aProportion of hospitalizations detected was calculated as the number of hospitalizations each surveillance system detected divided by the total number of hospitalizations estimated by the capture-recapture analyses.

The market share of each hospital remained consistent over the 4 seasons and was used to estimate hospitalization rates. The total estimated hospitalization rates based on capture-recapture estimates ranged from 23 per 10 000 persons in 2016–2017 to 40 per 10 000 in 2017–2018 (Table 1). Pediatric hospitalization rates were highest in the 2018–2019 season with 22 hospitalizations per 10 000 persons, followed by the 2019–2020 season with 8 hospitalizations per 10 000 persons (Table 1). Among specific age groups, the age group ≥65 years had the highest estimated hospitalization rate across all seasons, followed by the age group 50–64 years. The highest annual hospitalization rate was observed in the age group ≥65 years in 2017–2018 season with 170 hospitalizations per 10 000 persons.

DISCUSSION

By combining data from independent surveillance systems that operate concurrently in the same population, we obtained a comprehensive estimate of the burden of seasonal and age-specific influenza hospitalizations. The capture-recapture estimates were consistently higher for adults and children across all seasons than what each surveillance system independently identified.

Influenza surveillance is vital for monitoring viral activity, quantifying the burden of severe disease based on hospitalizations and deaths, characterizing circulating influenza strains, and monitoring the effectiveness of vaccination. When there are 2 or more existing surveillance systems available, capture-recapture methods can be used to estimate the influenza hospitalization burden in a comprehensive, efficient, and inexpensive way. We utilized data from established independent surveillance systems that operate concurrently in Middle Tennessee. EIP captures every hospitalized patient with a clinician-ordered positive influenza test result. HAIVEN and NVSN are active, labor- and time-intensive surveillance systems where eligible hospitalized patients are identified, approached, and consented to be tested, independent of clinician-driven testing. While EIP is designed to capture all hospitalized cases that test positive based on clinician-driven testing, the disadvantage is that patients who were not tested would be missed. In addition, available clinical tests used for influenza diagnosis may have variable performance characteristics. Furthermore, during periods of high influenza circulation, influenza diagnosis can be made clinically without laboratory testing. HAIVEN and NVSN potentially would capture some of these cases since those systems test regardless of clinician-driven testing, but with such a laborious approach and need to consent for enrollment, even these active surveillance systems would not catch every patient hospitalized with influenza. These differences in surveillance systems were reflected in our estimates of proportion of cases identified by each system, where EIP consistently found more hospitalizations than HAIVEN, possibly due to the need for consent and enrollment only occurring 5 days per week. In the 2019–2020 influenza season, EIP detected the highest proportion of cases for children and adults. However, for assessments of pediatric cases, EIP did not consistently detect a larger proportion of cases compared to NVSN. This may be due to inconsistent clinician-ordered influenza testing in children from year to year [13]. Additionally, CDC influenza management guidelines do not require confirmation of influenza to begin treatment in outpatient facilities, which may contribute to less clinician-ordered testing. Our study encompassed 4 consecutive influenza seasons; during that period RT-PCR and other nucleic antigen tests were more widely introduced into hospital laboratories for clinical use and likely led to more frequent testing over time, potentially increasing detections by EIP [14].

Influenza burden estimates (eg, capture-recapture estimates) generally align with specific influenza virus circulation across seasons. These estimates are also generally consistent with the estimates reported by the CDC. Influenza A(H3N2) was the predominant strain during the 2016–2017 and the 2017–2018 influenza seasons and was the most common strain in all age groups [9, 10]. Although the 2016–2017 influenza season was moderate [9], the 2017–2018 season had high levels of emergency department visits and increased hospitalization rates [10], which correlated with what was found in our capture-recapture estimates, with 2017–2018 showing the highest total hospitalization rate compared to the other 3 influenza seasons. The 2018–2019 season was a moderate influenza season with influenza A(H1N1) circulating throughout the first part of the season and influenza A(H3N2) predominantly circulating beginning in February. The 2018–2019 season reported less activity in adults than in the prior season [11], which was reflected in the hospitalization rates. The 2019–2020 season also had influenza A(H1N1) as the predominantly circulating virus; however, the season was shortened due to nonpharmaceutical interventions, such as implementing school closures, mask wearing, and handwashing, to control COVID-19 [5]. Our estimated hospitalization rates for the 2019–2020 season were similar to the estimated hospitalization rates for the previous season, both of which had influenza A(H1N1) circulating. Our estimated hospitalization rates for Middle Tennessee, which are similar to the national estimates reported by the CDC [15–18], followed the same age-specific trends with those aged ≥65 years having the highest hospitalization rates. The CDC estimates its hospitalization rates [19] from surveillance data collected through FluSurv-NET, which is based solely on EIP data [20]. Their reported hospitalization rates are adjusted for underdetection based on the proportion of patients tested for influenza and the average sensitivity of influenza testing. Similar to our approach, the CDC uses EIP data to model influenza burden; however, their estimates rely only on EIP. Our approach to use EIP data in conjunction to HAIVEN/NVSN provides an alternative method involving additional available surveillance systems.

Our study has several limitations. Capture-recapture estimates rely on the fulfillment of several assumptions. The estimates are valid under the assumptions that the probability of being captured by one system is independent of capture by the other system, the study population remains approximately constant, and the cases can be matched from both systems. Fulfillment of the independence assumption is difficult to determine when only 2 sources of data are available for analyses, as in our study. HAIVEN and NVSN actively recruit participants and conduct influenza testing regardless of physician driven testing, whereas EIP identifies every positive influenza case based on physician-ordered tests, so the assumption that these surveillance systems remain independent is difficult to verify. However, provided that these assumptions are met, using the capture-recapture method gives a better estimate of hospitalizations than using either surveillance system individually. We do not have information regarding the probability of consenting to participate in active surveillance based on clinician-driven testing. Nevertheless, HAIVEN and NVSN purposely enrolled based on fulfillment of a clinical case definition rather than clinician testing, and participants of HAIVEN and NVSN were not provided with the results of research laboratory testing. The capture-recapture estimates relied on data computed during the entire season, and while real-time estimates would be of interest, several specific steps would need to be taken to make that feasible, such as continuously updating and validating data to allow for individual-level results to be linked. While using 2 surveillance systems to calculate the burden of influenza-related hospitalizations provides a better estimate than estimates from each system alone for Middle Tennessee, these estimates may not be generalizable to other places in the country, as they may have experienced different influenza activity as usually influenza rates are highest in younger children compared to older children [13, 21]. We were unable to further stratify children into smaller age groups due to small sample size and were therefore unable to determine hospitalization rates by additional pediatric age groups. The proportion of hospitalizations that each surveillance captured may be specific to the included hospitals as testing practices may differ between healthcare systems [22] and higher market shares at participating hospitals would increase the precision of the estimates. While HAIVEN was conducted 5 days per week, which may be a potential limitation, the enrollment period includes admissions that had occurred in the previous nonactive enrollment days.

We used several surveillance systems and capture-recapture methods to obtain more comprehensive hospitalization estimates overall and by season and age group. Implementation of systemic testing can improve surveillance and allow for better representation of the disease burden as well as improve opportunities for infection prevention, control, and management. Increased testing for acute respiratory infections would be beneficial for surveillance systems in place, such as EIP. While using single existing surveillance systems as the only data source may lead to underestimation of disease burden, combining data from existing systems provides an inexpensive and data-driven alternative to better estimate the burden of influenza hospitalizations. By combining data from 2 surveillance systems, we estimated a larger burden of disease than that reported by each individual system. The significant underrecognized burden of influenza hospitalization underscores the importance of influenza prevention and supports health policies and efforts directed to increase influenza vaccination uptake and improvement of influenza vaccine effectiveness.

Supplementary Data

Supplementary materials are available at The Journal of Infectious Diseases online. Consisting of data provided by the authors to benefit the reader, the posted materials are not copyedited and are the sole responsibility of the authors, so questions or comments should be addressed to the corresponding author.

Notes

Acknowledgments. H. K. T. and C. G. G. conceptualized the study. H. K. T. acquired funding for the study. H. K. T. and N. H. acquired data. A. C. H. and Y. Z. conducted data analyses. A. C. H. prepared the first manuscript draft. A. C. H., Y. Z., D. W., T. M., J. D. C., N. H., C. H. T., S. M. O., J. F., S. G., W. S., C. G. G., and H. K. T. critically revised, edited, and approved the manuscript.

Disclaimer. The contents of this work are solely the responsibility of the authors and do not necessarily represent official views of the National Center for Advancing Translational Sciences, the National Institutes of Health, or the Centers for Disease Control and Prevention (CDC).

Financial support. This work was supported by the Centers for Disease Control and Prevention through an Emerging Infections Program cooperative agreement (grant number CK17-1701); by the CDC under cooperative agreement IP15-002, including U01IP000979; and by the National Center for Advancing Translational Sciences (Clinical and Translational Science Award number UL1 TR002243). C. G. G. was partially supported by the National Institutes of Health (award number K24AI148459).