Using inpatient electronic medical records to study influenza for pandemic preparedness

Abstract Background We assessed the ability to identify key data relevant to influenza and other respiratory virus surveillance in a large‐scale US‐based hospital electronic medical record (EMR) dataset using seasonal influenza as a use case. We describe characteristics and outcomes of hospitalized influenza cases across three seasons. Methods We identified patients with an influenza diagnosis between March 2017 and March 2020 in 140 US hospitals as part of the US FDA's Sentinel System. We calculated descriptive statistics on the presence of high‐risk conditions, influenza antiviral administrations, and severity endpoints. Results Among 5.1 million hospitalizations, we identified 29,520 hospitalizations with an influenza diagnosis; 64% were treated with an influenza antiviral within 2 days of admission, and 25% were treated >2 days after admission. Patients treated >2 days after admission had more comorbidities than patients treated within 2 days of admission. Patients never treated during hospitalization had more documentation of cardiovascular and other diseases than treated patients. We observed more severe endpoints in patients never treated (death = 3%, mechanical ventilation [MV] = 9%, intensive care unit [ICU] = 26%) or patients treated >2 days after admission (death = 2%, MV = 14%, ICU = 32%) than in patients treated earlier (treated on admission: death = 1%, MV = 5%, ICU = 23%, treated within 2 days of admission: death = 1%, MV = 7%, ICU = 27%). Conclusions We identified important trends in influenza severity related to treatment timing in a large inpatient dataset, laying the groundwork for the use of this and other inpatient EMR data for influenza and other respiratory virus surveillance.


| INTRODUCTION
The US Food and Drug Administration (FDA) has been exploring the feasibility of utilizing real world data, such as administrative claims and electronic medical records (EMR), to support decision making before or during a public health emergency. 1 Administrative claims provide information regarding exposure to medication dispensings and many outcomes, but use of these databases during a public health emergency can be limited by the time it takes for these databases to "settle." 2 In addition, there is often limited ability to assess detailed information regarding inpatient medication use and care received during hospitalization in claims databases. EMR data can provide timely and detailed clinical information, and the feasibility of conducting public health surveillance with EMRs has previously been demonstrated. [3][4][5][6][7][8][9] However, capturing and analyzing these data in real-time during a public health emergency are a challenge unless systems are already in place. Simonsen and colleagues have highlighted the use of seasonal influenza as an example infection when developing systems built on "big data" for infectious disease surveillance. 10 As part of the FDA's Sentinel System, 11,12 we explored the feasibility of utilizing inpatient EMR data for collecting and analyzing treatments and outcomes in hospitalized patients to support the FDA's need for timely information during a public health emergency. 13 Sentinel is an active surveillance system that uses routinely collected electronic healthcare data to support FDA's regulatory decision making. Our objective was to assess the ability to identify medications, severity, and other key data relevant to seasonal and pandemic respiratory virus activity in hospitalized patients, using seasonal influenza as a use case. We describe the baseline characteristics, healthcare utilization, complications, and endpoints of hospitalized adults with an influenza diagnosis from March 2017 through March 2020, using data on discharged patients.

| Data sources and study population
This was a retrospective descriptive study among adults hospitalized from March 1, 2017, through March 31, 2020. We leveraged an existing EMR dataset from HCA Healthcare for our study which includes 140 hospitals and is updated frequently to support a pragmatic trial at HCA Healthcare. 14,15 This Sentinel System study was a public health surveillance activity conducted under the authority of the FDA and, accordingly, was not subject to Institutional Review Board oversight. [16][17][18] March 2017 was selected as the study start as this is when the hospitals systematically began providing medication administration data. Please see Appendix S1for the study design diagram. Within this period, we identified hospitalizations with an influenza diagnosis (via International Classification of Diseases, Tenth Revision, Clinical Modification codes [ICD-10-CM]; see Appendix S2 for code lists). We included data on patients discharged and with complete billing only and did not include hospitalizations for patients still admitted or not completely coded when the datasets were created. For reference, we also examined characteristics of all hospitalizations captured in the database.

| Demographics, high-risk conditions, and treatments
We assessed demographics (age, sex, and race) on the date of hospital admission. We examined conditions that may increase a person's risk of serious complications from influenza, including chronic respiratory disease (asthma, chronic obstructive pulmonary disease [COPD], and other chronic respiratory disorders), chronic cardiovascular disease, liver or renal disorders, immune disorders, diabetes, obesity, hematological disorders, and smoking. These were identified via ICD-10-CM diagnosis codes documented throughout the hospitalization. We also assessed pregnancy status via diagnosis and procedure codes that are markers for pregnancy as well as those for gestational age recorded at any point during the hospitalization (see Appendix S2 for code lists.) We identified influenza antiviral treatment administration (or lack thereof) on the day of admission, within 2 days of admission, and beyond 2 days of admission (oseltamivir, zanamivir, peramivir, baloxavir). We also examined antibiotic use during hospitalization. We used brand names, generic names, National Drug Codes (NDC), and Healthcare Common Procedure Coding System (HCPCS) procedure codes to define treatments. Please see Appendix S3 for the medication search strategy and relevant procedure codes.

| Complications
We looked for complications coded during the hospital stay and examined those complications by whether they were coded as present on admission or after admission. Complications included pulmonary complications, inflammatory conditions, myocardial infarction, stroke, and sepsis (see Appendix S2 for code lists). We also examined death in the hospital (i.e., discharged expired).
To describe markers of illness severity we examined intensive care unit (ICU) stays, use of supplemental oxygen, bilevel positive airway pressure (BiPAP), mechanical ventilation (MV), and extracorporeal membrane oxygenation (ECMO) (see Appendix S2 for code lists).
Length of stay was calculated based on admission and discharge dates.

| Ordinal endpoints
We assessed select complications that are associated with severe influenza, and other respiratory infections, and may be of interest in future studies of MCM safety and effectiveness as ordinal endpoints.
Ordinal endpoints were as follows: (1) Table 1 summarizes the number of hospitalizations with an influenza diagnosis, demographic characteristics, and high-risk conditions captured during the hospital stay stratified by timing of influenza antiviral treatment. Among the influenza hospitalizations, there were more females than males (58%), and more than half of the patients were over the age of 65 years (56%) on admission. Race information was well captured with just slightly over 1% of hospitalizations missing race information.

Similar proportions of smokers were captured in hospitalizations with
influenza diagnoses compared with all-cause hospitalizations (18% vs. 18%). Approximately 3% of patients with an influenza diagnosis had evidence of pregnancy. Antiviral treatment administrations were recorded in 85% of the influenza hospitalizations; 64% were treated on the admission date or ≤2 days after admission. Patients treated later, >2 days after admission, were older and generally had more high-risk conditions than those treated earlier. Patients never treated with an influenza antiviral were younger and certain high-risk conditions such as obesity, ischemic heart disease, heart failure, liver and renal failure, hematological disorders, and smoking were documented more frequently than patients treated early in their hospitalization. Influenza was the principal diagnosis in 52% of hospitalizations treated on the day of admission, 43% treated ≤2 days of admission, 29% treated >2 days of admission, and 16% of hospitalizations that had no evidence of treatment during their stay.  we examined administration of any antibiotic during the hospitalization and found 77% of hospitalizations with influenza diagnosis also had evidence of an antibiotic administration. For oxygen delivery, we were able to identify MV (6% on admission, 6% after admission) and

| MCMs and oxygen delivery in influenza hospitalizations
supplemental oxygen use (32% on admission, 51% after admission).
We did not identify any BiPAP use with procedure codes and identified three hospitalizations with an influenza diagnosis and ECMO during their stay (Table 2).      Note: Includes discharges for final billed patients only. The analysis does not include inpatient stays for patients still admitted or not completely coded by the data pull date. a Any complication includes: Pneumonia associated with influenza, acute respiratory failure, chronic respiratory failure, ARDS, supplemental oxygen, BiPAP, mechanical ventilation, ECMO, inflammation of heart, brain, or muscle tissue, myocardial infarction, ischemic stroke, sepsis, ICU stay, or death. b Severe complication includes the following: pneumonia associated with influenza, mechanical ventilation, ICU stay, or death. c ICU defined via revenue codes. advocated as useful for evaluating treatment of hospitalized influenza patients [19][20][21] and during the COVID-19 pandemic, 22 we explored the feasibility of capturing these endpoints and reported unadjusted rates to inform future studies. We found that using coded data to capture severe endpoints such as in-hospital death, ICU stays, and MV during influenza hospitalizations was feasible.

| Complications and ordinal endpoints in influenza hospitalizations
We found the majority (85%) of hospitalizations with an influenza diagnosis had record of an antiviral treatment administration during their stay, and 64% had evidence of treatment within 2 days of admission (29% on admission date, 35% ≤ 2 days following admission).
Patients treated >2 days after admission had more comorbidities than patients treated earlier. Similar to another recent study, severe endpoints were lowest among those treated on admission, and highest among patients treated >2 days after admission or not treated during their stay. 23 The exception was death, which occurred most frequently in patients not treated during their stay.
There was no evidence of influenza antiviral administration in 15% of hospitalizations with influenza diagnosis, and we observed frequent documentation of cardiovascular conditions (e.g., ischemic heart disease and heart failure), obesity, and smoking among these patients.
In addition, just 16% of hospitalizations without antiviral administrations had influenza as the principal diagnosis code. It is possible some patients without evidence of antiviral treatment during hospitalization did not truly have influenza, and influenza was a differential diagnosis.
However, we were unable to confirm this hypothesis due to the lack of available influenza testing data at the time of this analysis.
The FDA plays a key role in ensuring access to safe and effective medical countermeasures (MCMs; e.g., diagnostic and treatments) during a public health emergency. 1 Information about MCM safety and effectiveness becomes even more important when an investigational MCM is made available during an emergency. However, capturing and analyzing real-time information during an emergency remains a challenge. Our study established the capacity for these inpatient EMR data to be used in an emergency while also providing important information about seasonal influenza for future work.
In our study, we were able to capture influenza antiviral treatments along with administration dates and times. This bodes well for future studies using inpatient EMR data to examine medications administered in the hospital. However, it is important that future studies explore medications of interest within their data source. Understanding how medications are captured in data sources used for future studies, and recognizing situations when they may not be completely captured, especially within specific care settings (e.g., intra-operatively administered medications) 24 is not a challenge unique to this study. Considerations for ensuring real-world data are fit for purpose have been commented on previously. 2,25 We examined oxygen delivery as well as ordinal endpoints in hospitalizations with influenza diagnosis codes. Although we found that up to 40% of influenza non-ICU hospitalizations had evidence of supplemental oxygen use, we understand that oxygen use may be underestimated if only procedure codes are relied upon. 26 We were unable to identify BiPAP in this study, which was not unexpected, as billing practices may bundle this with other care and our study identified oxygen delivery based on diagnosis and procedure codes. We expect that both oxygen supplementation and BiPAP use are included in nursing documentation within many EMR systems, and thus it may be possible to extract such information as needed. Future studies in similar inpatient datasets should consider exploring the feasibility of retrieving and analyzing nursing documentation to examine the capture of and ability to attain more specific information regarding type and duration of oxygen therapy.

| Strengths and limitations
The major strength of this study was the size of the data source. We We restricted our analysis to only discharged patients with complete billing information. While this means the data were not as "fresh" as possible, it also means the data are complete. Others have asserted that during an evolving public health emergency, information used for decision making should be stable and complete. 2 There are several details to consider when interpreting our study results. We were unable to examine patient characteristics, medication use, or care delivered before or after the hospitalization and relied on conditions coded during hospitalization to examine baseline and high-risk conditions. We used diagnosis and procedure codes to examine conditions and procedures and used revenue codes to define ICU stays. We did not have access to laboratory results to confirm influenza diagnosis at the time of this study, although such data are available in this source. Illness onset dates were also not available, a limitation that will be common in claims data sources as well as other