With fluctuating vaccine supplies, variable onset and severity of influenza seasons each year, and new recommendations for use of influenza vaccine in children, an accurate, informative influenza surveillance system is greatly needed. During the 2003–2004 influenza season, analysis of data from 2 independent surveillance systems, both of which included children <5 years of age, provided better estimates of hospitalization rates since it accounted for those cases undetected by each system.
NVSN attempted to enroll all potential influenza admissions on surveillance days and used the most sensitive and specific diagnostic tests to detect influenza (18
). Reliance on viral culture alone for influenza diagnosis would have missed 21% of NVSN cases, whereas use of RT-PCR would have missed only 7%. The combination of these techniques increased the detection of influenza by the NVSN. In addition, nonsurveillance days and children whose parents refused enrollment were taken into account in NVSN rate calculations. With intense surveillance, NVSN detected 73% of influenza hospitalizations estimated by the capture-recapture analysis. Exclusion of the 4 possible nosocomial cases increased the sensitivity of NVSN to 84%. NVSN selection criteria were established to specifically exclude nosocomial cases, including children discharged within 4 days of readmission. One child was enrolled by NVSN and tested negative for influenza virus on admission but had a clinical laboratory test result indicating influenza after >1 week of hospitalization. Three other children were excluded by NVSN criteria because of a recent hospitalization. However, with available information, whether these were nosocomial infections could not be determined. Based on results of the capture-recapture analysis, NVSN modified its methodology in subsequent years to include children recently hospitalized.
The EIP surveillance system sought to find all hospitalized children with positive clinical laboratory test results for influenza. One limitation of the EIP was that influenza ascertainment relied on a diagnostic test ordered by the physician. Another limitation was the sensitivity of the rapid influenza detection tests. When tests were not ordered or yielded false-negative results, influenza cases were undetected. EIP surveillance was cheaper and logistically simpler to implement than NSVN. Since EIP was considered a public health response program in Tennessee, it did not require parental informed consent. However, EIP will underestimate the impact unless combined with additional information on the proportion of patients with true cases who are tested and the sensitivity of the diagnostic tests used. During its first year of influenza surveillance in Davidson County, EIP missed only 1 patient who could have potentially been detected. However, because the rapid influenza antigen test, the only clinical laboratory influenza test used in these patients, was less sensitive than RT-PCR plus viral culture, the estimated sensitivity of EIP for children who actually had clinical laboratory tests performed was 64%. When compared with viral culture, these tests have a sensitivity ranging from 44% to 95% and a specificity ranging from 76% to 100% (27–30
). Although the ability of EIP to detect influenza cases was dependent on these test characteristics, the primary reason for EIP's underestimation of rates was that diagnostic tests for influenza were not ordered for most children admitted with influenza. The capture-recapture analysis indicated that only 38% of children <5 years of age hospitalized with influenza were correctly identified by routine diagnostic tests. Thus, not detecting influenza during hospitalization resulted not only in underestimating the impact of influenza, but also in providing limited opportunity for appropriate antiviral therapy.
Capture-recapture methods emerged as an adaptation of techniques used by wildlife researchers to obtain better counts of difficult-to-enumerate wild animals. The simplest technique uses 2 samples or lists. Using the number of individuals caught in each sample (captures) and the number of subjects from the first sample that were captured again by the second sample (recaptures), one can estimate the number of subjects not caught in either sample, thus providing an estimate of the total population size (31–34
). The estimation directly accounts for different capture probabilities of each sample, and allows one to obtain estimates using 1 source that operated 4 days a week (NVSN) and the other that operated continuously (EIP).
Since being identified in 1 system did not influence the possibility of identification in the other system, the independence of the 2 systems was assumed. The independence assumption could have been violated if some factor, such as severity of influenza illness or viral load, varied among subjects and the likelihood of detection increased in both systems with increasing severity or viral load. In this case, the Peterson method would underestimate the true population size. In addition, both systems would likely miss children with very low or no influenza viral loads, such as those admitted late in the course of illness. This would also underestimate the true rates.
No significant migration occurred in Davidson County during the study, and the study population was restricted to county residents and assumed to be closed. This study was conducted during a single influenza season and there were relatively small numbers of cases identified, which precluded detailed subgroup analyses. However, the final estimation of influenza hospitalization rates was consistent with previous reports of the 2003–2004 influenza season and with previous research indicating that children <24 months of age have hospitalization rates similar to those of persons >65 years of age (11,12,20,35
). This estimation also highlights the great impact of influenza, particularly in children <6 months of age during a moderately severe influenza season. Current vaccines are poorly immunogenic in this age group and have not been approved for these children. Thus, vaccination of household contacts and out-of-home caregivers of children <6 months of age is recommended. Additional influenza vaccination of children 6–23 months of age has also been recommended to limit their exposure (35,36
). As immunization rates in families of young children increase and routine vaccination for children 6–23 months of age is implemented, surveillance systems must be in place to effectively measure the impact of these preventive strategies.
Combined NVSN and EIP systems analyzed with the capture-recapture approach appear well suited to this important task. EIP is a simpler and cheaper system for identifying children with influenza. Although EIP could estimate rates more accurately by adjusting for known sensitivities and specificities of clinical diagnostic tests, without information on the frequency of diagnostic testing, it would be impossible to determine and adjust for the proportion of influenza this system captures. Thus, the degree of underascertainment would be unknown. In addition, such diagnostic testing will likely change over time, making year-to-year comparisons of disease impact difficult. NVSN attempted to estimate the true impact of influenza hospitalizations by testing all children with specific admission criteria, adjusting for nonenrollment and nonsurveillance days, and providing an unbiased sample of influenza-positive children for further analyses such as vaccine effectiveness estimates. However, this system also underestimated the total influenza impact. The combined systems gave the best estimate of disease impact.
Currently, no population-based surveillance systems are available to monitor the influenza vaccine program in adults. Using a combination of 2 systems similar to NVSN and EIP could be a model for surveillance of influenza in adults. The more expensive and labor-intensive NVSN-type surveillance could be conducted at representative hospitals in a geographic area for limited periods during the influenza season (e.g., 1 day/week at each hospital). The EIP-type surveillance system could attempt to identify all persons admitted with influenza identified through routine testing. Capture-recapture methods could be used to more accurately estimate serious influenza impact. Comparison of patients could determine whether those identified through cheaper EIP methods were representative of all patients with respect to important characteristics such as influenza vaccination status and severity of disease. Capture-recapture techniques should be considered as methods to best use limited resources for essential surveillance activities.