Evaluating oseltamivir prescriptions in Centers for Medicare and Medicaid Services medical claims records as an indicator of seasonal influenza in the United States

Background Over 34 million residents of the United States aged 65 years and older are also Medicare prescription drug beneficiaries. Medical claims records for this age group potentially provide a wealth of data for better understanding influenza epidemiology. Objective The purpose of this study was to evaluate data on oseltamivir dispensing extracted from medical claims records as an indicator of influenza activity in the United States for the 2010‐11 through 2014‐15 influenza seasons. Methods We used Centers for Medicare and Medicaid Services (CMS) medical claims data to evaluate the weekly number of therapeutic oseltamivir prescriptions dispensed following a rapid influenza diagnostic test among beneficiaries 65 years old and older as an indicator of influenza timing and intensity. We compared the temporal changes in this indicator to changes in the proportion of influenza‐like illnesses among outpatient visits in the US Outpatient Influenza‐like Illness Surveillance Network (ILINet) by administrative regions defined by the US Department of Health and Human Services. Using the moving epidemic method, we determined intensity thresholds and categorized the severity of seasons for both CMS and ILINet data. Results Centers for Medicare and Medicaid Services oseltamivir data and ILINet data were strongly correlated by administrative region (median Spearman's ρ = 0.78; interquartile range = 0.73‐0.80). CMS oseltamivir data and ILINet data substantially agreed (Cohen's weighted κ = 0.62) as to the seasonal severity across administrative regions. Conclusions Our results support the use of oseltamivir dispensing in medical claims data as an indicator of US influenza activity.


| INTRODUC TI ON
National and international surveillance for influenza informs important decisions, such as selecting vaccine components. 1 In the United States, national surveillance for influenza has multiple components including syndromic surveillance for influenza-like illness (ILI) in an outpatient setting and laboratory-based surveillance for detecting influenza viruses from clinical specimens. 2 Other components of national surveillance include inpatient sentinel surveillance for laboratory-confirmed influenza, surveillance for deaths attributable to pneumonia and influenza, and surveillance for influenzaassociated pediatric mortality. [2][3][4] Additional data sources-such as medical records, survey data, and Internet search queries-improve context when interpreting statistics from national surveillance. 5 Medical claims data provide valuable insight into care provided to patients covered by the payer. Previous research found a strong correlation between national surveillance for ILI in the United States and prescriptions for antivirals with activity against influenza. 6 Medicare claims data are particularly attractive for analysis of influenza in the United States, as patients 65 years old and older are a large population at risk of influenza infection, and over 34 million people in this age group are Medicare prescription drug beneficiaries. 7,8 However, these data only include information needed for billing and do not include results from diagnostic assays. Previous research captured trends in ILI surveillance at national and regional levels using syndromic data on outpatient visit diagnosis codes in medical claims data, but these results could represent illnesses from a number of respiratory pathogens. 9 While developing an outcome to detect influenza-associated clinic visits in a study assessing the relative effectiveness of high-dose vs standard-dose influenza vaccines, we noted temporal associations between Medicare claims for oseltamivir prescriptions and national surveillance for influenza viruses in clinical specimens. 10 Here, we investigate whether trends in oseltamivir prescriptions dispensed to Medicare beneficiaries are temporally and spatially associated with trends in outpatient ILI and trends in detection of influenza viruses in clinical specimens

| ME THODS
The Centers for Medicare and Medicaid Services (CMS) compiles billing claims for services rendered to beneficiaries by healthcare providers. We linked claims data for Medicare Part B and Part D, which pay for community-level care and prescription drugs, respectively. Community-level care includes physician services and other non-inpatient services such as durable medical equipment, laboratory services, and imaging services. From January 3, 2010, until October 31, 2015, we compiled both the weekly number of therapeutic oseltamivir prescriptions of 75 mg twice daily for 5 days among beneficiaries 65 years old and older and the subset dispensed within 2 days of a rapid influenza diagnostic test (RIDT). Therefore, we excluded prophylactic prescriptions of oseltamivir at a once daily rate. We used oseltamivir dispensed within 2 days of a RIDT because this suggests the clinician knew the result of the RIDT when prescribing oseltamivir. 11 We stratified the data by administrative region as defined by the Department of Health and Human Services.
We defined week 1 as the first week with at least 4 days in a year.
Each influenza season started on week 40 and ended the following year on week 39. We evaluated the weekly number of oseltamivir prescriptions-with and without a RIDT-as a potential indicator of influenza activity.
To assess seasonal severity, we applied the moving epidemic method (MEM) developed by Vega and others. 12 These researchers applied the MEM to data for ILI and acute respiratory illnesses in 28 European countries to demonstrate the method's value in normalizing disparate data. 13 Additionally, the researchers applied the MEM to regional surveillance networks within Spain to demonstrate the method's value in early detection of influenza. 14 One component of the MEM defines epidemic periods for each season, an interval when influenza activity is high. Another component of the MEM is constructing intensity thresholds (ITs) using confidence intervals.
We applied the MEM to our data as follows: For each region, we found the 6 largest weekly counts from each of the 5 epidemic periods from the 2010-11 through 2014-15 seasons. For each region, we used these 30 counts to construct 1-sided 100% × (1α) confidence intervals for the geometric mean assuming a lognormal distribution at α = 0.50, 0.10, and 0.02. We labeled the upper limits of these confidence intervals as IT 50 , IT 90 , and IT 98 . We categorized each season as a low-severity season when the largest weekly value falls below IT 50 , a moderate-severity season when between IT 50 and IT 90 , a high-severity season when between IT 90 and IT 98 , and a very high-severity season when above IT 98 .
To assess the validity of using oseltamivir prescriptions among beneficiaries to measure seasonal severity, we compared results from this MEM analysis to analogous results using national surveillance data compiled by the Centers for Disease Control and Prevention (CDC). We used the weighted proportion of ILI among outpatient healthcare providers participating in the US Outpatient Influenza-like Illness Surveillance Network (ILINet) as the primary reference. As secondary references, we used (i) the proportion of specimens testing positive for influenza virus among those specimens submitted to the World Health Organization Collaborating Laboratories and the National Enteric Virus Surveillance System in the United States (denoted influenza virus data hereafter) for influenza testing and (ii) the Goldstein index, a proxy for the weekly proportion of laboratory-confirmed influenza infections among those seeking care in the outpatient setting. 15 These surveillance systems are described elsewhere. 2 We did not limit surveillance data to people 65 years old and older, as these data are not available stratified by both region and age group. Using the results of the MEM analyses, we used Cohen's weighted κ to measure agreement between the severity of each season as categorized with the CMS data and the severity of each season as categorized with the national surveillance data. 16 We interpreted Cohen's weighted κ using the Landis-Koch classification. 17 As a sensitivity analysis to assess our implicit assumption of a steady population at risk in the CMS data, we also considered the weekly proportion of beneficiaries meeting the case definition (cases per person per week).
We used Spearman's ρ to measure the association between weekly counts in the CMS data and the weekly indicators in the national surveillance data. To assess the relative timing between the CMS data and the national surveillance data, we used time-lagged cross-correlations of the time series from the CMS data and national surveillance data. 18     The ITs for the number of oseltamivir prescriptions varied across regions (Table S1). Agreement between the seasonal severity using the weekly number of prescriptions and the weekly proportion of ILI was substantial (κ = 0.62; Table 3). The 2010-11 season was categorized as a low-severity season using the weekly number of prescriptions for all regions, but most regional severities were higher when using national surveillance data ( Figure 2).  Table 2).

| Oseltamivir dispensed following a RIDT
The median difference between the week with the largest number of prescriptions following a RIDT within a region and the larg- Within regions, ITs for the weekly number of oseltamivir prescriptions following a RIDT varied (Table S1). Agreement between seasonal severity using the weekly number of prescriptions and seasonal severity using the weekly proportion of ILI was substantial (κ = 0.62; Table 3). When the CMS data and data from another surveillance system exceeded an IT, the time difference was within 3 weeks 85% of the time. Outliers to this trend included a late peak in the oseltamivir data for region 1 and region 2 during the 2013-14 season relative to influenza virus data ( Figure 5).

| Comparison of oseltamivir prescriptions with oseltamivir prescriptions following RIDT
The national weekly number of oseltamivir prescriptions strongly correlated with the weekly number of prescriptions following a RIDT (ρ = 0.98). Similarly, the regional weekly number of prescriptions strongly correlated with the weekly number of prescriptions following a RIDT (median ρ = 0.90, IQR = 0.87-0.93). The proportion of oseltamivir prescriptions following a RIDT among those receiving prescriptions was seasonal ( Figure 6) and varied by region (Table 1).

| D ISCUSS I ON
The weekly number of oseltamivir prescriptions dispensed to Medicare beneficiaries 65 years old and older within 2 days of a RIDT strongly correlated with clinical and influenza virus data for all ages from existing national surveillance systems for influenza ( Table 2). We found no evidence of a meaningful time lag between prescriptions following a RIDT with outpatient ILI, indicating that the 2 sources of data yield similar conclusions about the timing of influenza activity. At the national level, our indicators of influenza assigned the same seasonal severity categories as previous work with analogous methods using national surveillance data for people 65 years old and older. 20 Our results suggest that oseltamivir prescriptions following a RIDT may serve as a proxy for laboratory-confirmed influenza in epidemiological studies using medical claims in the United States.
The peak week of the number of oseltamivir prescriptions following a RIDT was often different from the peak week of ILI, especially for the 2011-12 season ( Figure 5). We did not find as many differences when we compared the weekly number of prescriptions following a RIDT with the Goldstein index, suggesting these differences in peak timing may be partly attributable to ILIs with a non-influenza etiology ( Figure 5). The late peaks of prescriptions following a RIDT during the 2013-14 season in the Northeast (regions 1 and 2) may be attributable to a second wave of influenza ( Figure 5). The first wave was predominantly influenza A (H1). 21 The second wave consisted of both influenza A (H3) and influenza B, and this wave was more prominent in people 65 years old and older. 22 Also, the second wave was more prominent in the Northeast relative to the other US regions. 22 Given this context, these late peaks suggest our indicator of influenza is sensitive to incidence of influenza among people 65 years old and older.
Our to the community setting, as the medical claims data do not specify drugs administered during inpatient care. Conversely, influenza illnesses that do not come to medical attention are also absent from medical claims data. Therefore, the pathogenicity and virulence of circulating influenza viruses-which may differ between people 65 years old and older and younger people-affect our indicator in complex ways. 25,26 Results from the MEM analyses of medical claims data agreed substantially with results from ILINet; however, they differed with respect to results from the influenza virus data (Table 3).
Agreement in seasonal severity was poor in the 2010-11 season ( Figure 2), a season that nationally was less severe for people 65 years old and older relative to younger people. 20 In general, using ILI as a case definition may have low predictive value for influenza among young children because of high incidence of respiratory syncytial virus, rhinovirus, and metapneumovirus infections in this age group, especially when influenza is not circulating widely. [27][28][29] This disagreement in assessment of seasonal severity may be attributed to the discrepancy in source populations, suggesting our indicators are specific to influenza activity in people 65 years old and older.
The use of RIDT and oseltamivir in community settings varies within and across seasons. 11,30 Reliance on RIDT in the community TA B L E 3 Cohen's weighted κ measuring agreement between seasonal severity in the medical claims data and seasonal severity in the national surveillance data from the 2010-11 through 2014-15 influenza season

ACK N OWLED G EM ENTS
Our results rely on data contributed by many healthcare providers, public health practitioners, and laboratorians, to whom we are grateful for the privilege of working with these data. We thank Lynnette Brammer and Jerome Tokars for their careful review.

D I SCL A I M ER
The findings and conclusions in this report are those of the authors and do not necessarily represent the official position of the