Incidence and Outcomes of Non–Ventilator-Associated Hospital-Acquired Pneumonia in 284 US Hospitals Using Electronic Surveillance Criteria

This cohort study examines the incidence, variability, outcomes, and population attributable mortality of non–ventilator-associated hospital-acquired pneumonia in the US.


Introduction
Hospital-acquired pneumonia is the most common health care-associated infection in the United States and is associated with high morbidity, mortality, and health care use. 1,2Most cases occur in nonventilated patients.Crude mortality rates for non-ventilator-associated hospital-acquired pneumonia (NV-HAP) are similar to those for ventilator-associated pneumonia (VAP). 1,3,4However, most hospitals only have surveillance and prevention programs for VAP but not for NV-HAP.
Hospitals' limited attention to NV-HAP is partly due to difficulty defining and tracking NV-HAP and a limited understanding of its burden and preventability.The surveillance definitions published by the US Centers for Disease Control and Prevention National Healthcare Safety Network (CDC-NHSN) 5 include many subjective and ambiguous criteria that are complicated and difficult to apply in a clear and consistent manner, even for experienced clinicians. 6,7These include changes in oxygenation, the quality and quantity of respiratory secretions, and interpreting chest radiographs.
These criteria are prone to high levels of human error and interobserver variability and correspond inconsistently with histological pneumonia. 8,9Surveillance using discharge diagnosis codes is similarly limited because of their poor sensitivity, specificity, and variability within and between hospitals in how and when codes are applied. 10,11These challenges deter hospitals from developing NV-HAP surveillance and prevention programs, complicate the assessment of NV-HAP prevention initiatives, and make it difficult to estimate the national burden of NV-HAP.
Electronic health record (EHR) systems allow for the possibility of identifying NV-HAP using detailed electronic clinical data that can be applied in a consistent, automatable, and efficient manner across hospitals instead of existing approaches based on manual medical record reviews or administrative coding. 12This approach does not overcome the limited accuracy of traditional surveillance definitions but does have the potential to generate comparable estimates of incidence and outcomes between hospitals and across time in an efficient and reproducible manner.
We implemented a previously developed and validated electronic NV-HAP surveillance definition 12,13 predicated upon identifying patients with new and sustained deteriorations in oxygenation, abnormal temperature or white blood cell count, chest imaging, and new antibiotic starts.The electronic surveillance definition was applied to EHR data from a large and diverse set of US hospitals to estimate the incidence and variability of NV-HAP and its association with outcomes.
We then used detailed daily clinical data to estimate the attributable population mortality of NV-HAP, accounting for both baseline confounding and time-dependent bias.

Methods
This retrospective cohort study was approved by the institutional review boards of the Veterans Affairs (VA), University of Utah, and Harvard Pilgrim Health Care Institute.Informed consent was waived because consent would not be feasible and the study posed minimal patient risk.This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.

Design, Data Source, and Study Populations
[17] We identified all hospitalizations for adults ages 18 years or older admitted to VA acute care facilities between January 1, 2015, and November 30, 2020, and HCA Healthcare facilities between October 1, 2018, and May 31, 2020.For descriptive analyses, we included all hospitalizations.For the weighted analysis estimating attributable mortality risk, we only included hospitalizations of at least 3 days with complete data for the first or second day of admission for the following time-varying variables: hospital service, white blood cell count, hematocrit, platelets, sodium, glucose, and creatinine.

Electronic NV-HAP Surveillance Definition
The electronic NV-HAP surveillance definition was designed to mirror traditional surveillance definitions. 12The definition requires a decrease in oxygen saturation or increase in supplemental oxygen sustained for 2 or more days after 2 or more days of stable or improving oxygenation, plus an abnormal temperature (Յ36 °C or Ն38 °C) or white blood cell count (<4000 or Ն12 000 cells/mm 3 ), plus completion of chest imaging (x-ray or computed tomography), plus administration of 3 or more days of new antimicrobials starting on the first or secondary day of oxygen deterioration. 12Previous studies suggest this definition performs similarly to traditional CDC-NHSN criteria in terms of incidence, mortality, and clinical correlates. 12,18Further details regarding these criteria and the SAS code to apply the criteria are available in GitHub and eAppendix 1 in Supplement 1. 19 We extracted patients' demographics, comorbidities, vital signs, supplementary oxygen devices, laboratory tests, and discharge diagnosis codes.We defined comorbid conditions (Table 1) using the Elixhauser method applied to the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) 20 and the International Statistical Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) 21 codes via software developed by the Agency for Healthcare Research and Quality. 22,23We calculated a summary Elixhauser index score using the method of van Walraven. 24We extracted the first, minimum, maximum, mean, and median vital signs, pulse oximetry, supplemental oxygen use, and laboratory results for each calendar day of hospitalization.We identified daily specialty service and ward type (medical/surgical ward or intensive care unit).We also extracted facility characteristics, including bed size, geographic region, and teaching status, defined as the presence of graduate medical residents or medical students.
Race data were collected according to the data entered into the EHR.The majority of facilities collect race as self-reported.Race was used as a time-fixed covariate since it is a marker for additional baseline risk.Ethnicity was not collected because it was not consistently available.Categories of race included were Asian, Black, White, and other.For HCA, the other category included American Indian or Alaskan Native, Hawaiian or Pacific Islander, Hispanic, Multiracial, other, unknown, and for VA, other included American Indian or Alaska Native, Native Hawaiian or other Pacific Islander, mixed, or missing.This is described in Table 1.

Medical Record Review
The medical records of 250 randomly selected VA hospitalizations meeting electronic NV-HAP criteria were reviewed by 3 physicians (B.E.J., S.E.S., M.A.C.).Each record was independently reviewed by 2 physicians to determine: (1) whether the bedside clinical team documented NV-HAP; (2) whether the event met CDC-NHSN PNU criteria for NV-HAP 5 ; (3) whether NV-HAP was recorded in the discharge summary; and (4) whether the reviewer judged the case to be consistent with NV-HAP.Details, including the medical record review guide with specific definitions, are available in eAppendix 2 in Supplement 1 and separate work. 13

Statistical Analysis Descriptive Analysis
Incidence of NV-HAP (events per 100 hospitalizations and 1000 patient-days), patient characteristics, and inpatient mortality were calculated.To visualize facility-level variation, we plotted the incidence of NV-HAP surveillance events per 100 hospitalizations for each facility, ranked by incidence.Among the cases undergoing medical record review, we calculated the positive predictive value of electronic NV-HAP surveillance criteria vs bedside clinicians, CDC/NHSN criteria, and reviewers and assessed interrater agreement using Cohen κ.

Estimating the Attributable Mortality of NV-HAP
We estimated the attributable mortality of NV-HAP by modeling what could happen to population mortality if we could eliminate NV-HAP from the population theoretically.This is not to say that we believe that NV-HAP can be entirely eliminated; rather, this is a statistical approach for estimating attributable mortality that effectively handles time-dependent bias. 25This approach can account for complex time-varying confounding that cannot adequately be addressed using standard regression procedures and that risk spurious associations. 26We modeled the cumulative risk of the primary outcome (inpatient death by 60 days follow-up) vs alive discharge (a competing risk for in-hospital mortality) for the entire hospital population with NV-HAP cases included (current care).We then modeled cumulative incidences of inpatient death by 60 days under the theoretical condition of eliminating all NV-HAP cases from the hospital population.We compared these values to estimate the population attributable risk of inpatient death that is associated with NV-HAP.
We used inverse probability weighting 27 based on an estimated propensity score on each day to adjust for the following time-fixed confounders: hospital characteristics (size, teaching status, and region), patients' demographics (age, race, or sex), comorbidities, and hospitalization in the preceding 90 days.We also adjusted for the following time-varying confounders, corresponding to the most recent measured value 2 days prior: hospital service, intensive care unit (ICU) status, and markers of illness severity, including oxygen delivery device, pulse oximetry, routine laboratory test results (hematocrit, platelets, sodium, glucose, creatinine), and days since any current routine lab measure.We also adjusted for categories of nonroutine laboratory measures (alanine aminotransferase, total bilirubin, and albumin) as detailed in eAppendix 3 in Supplement 1.The statistical code is available via GitHub. 19Missing pulse oximetry or oxygen supplementation were treated as normal.
We used a weighted Aalen-Johansen estimator of the cumulative incidence of inpatient death and discharged alive under (1) current care and (2) hypothetical elimination of NV-HAP. 28Daily propensity scores for an NV-HAP event were used to construct weighted time-varying estimates of the hazards for each competing event, which were then used to calculate hazards for each outcome. 27Statistical methods are described in full in eAppendix 3 in Supplement 1. Risk ratios and risk differences were calculated at 60 days follow-up by taking the ratio and difference, respectively, of the risk estimate with hypothetical elimination of NV-HAP vs current care.We generated 95% CIs using a nonparametric bootstrap to resample hospitalizations with replacement 500 times.
We conducted exploratory analyses stratified by age (ie, Յ65 years or >65 years), service group (medical, surgical, cardiology, neuroscience, oncology, and other), ICU status on hospital day 3, Elixhauser comorbidity index quartile, hospital number of beds, region, and hospital teaching status.
We also conducted a sensitivity analysis using inpatient death vs alive discharge by 30 days of follow-up as a secondary outcome.Hospital-level point estimates and bootstraps were computed separately for VA and HCA in order to maintain data security and then combined using weighted

Estimation of Attributable Mortality and Exploratory Analyses
Among the 4 038 974 hospitalizations included in the weighted analysis, the estimated cumulative risk of inpatient death for all patients by 60 days of follow-up was 1.87% under current care.Under the analysis corresponding to hypothetical elimination of NV-HAP, the estimated risk was 1.73% (absolute risk difference 0.14% [95% CI, 0.13%-0.14%];relative risk ratio, 0.927 [95% CI, 0.92.5-0.928])(Figure 3 and eAppendix 4 in Supplement 1).Differences in the cumulative risk of inpatient death vs alive discharge within 60 days were similar across stratifications by age, comorbidity burden, ICU status, service group, hospital size, teaching affiliation, and region, and when follow-up was limited to 30 days (eAppendix 4 in Supplement 1).

Discussion
Among 284 diverse US hospitals from 2 large health care networks, approximately 1 in 200 admissions met electronic surveillance criteria for a possible NV-HAP event, of whom 22% died in-hospital, another 8% were discharged to hospice, and only 38% were discharged directly home.
Incidence rates varied by the hospital from 1 per 300 admissions in the lowest quartile to 1 per 150 admissions in the highest quartile.We estimated that NV-HAP could be associated with up to 7.3% of hospital deaths.The incidence and crude individual mortality of NV-HAP was within the range reported in previous studies using point prevalence, 1 manual, 30 or semi-automated approaches. 31However, this study is the first to generate population-level estimates of the incidence, variation, outcomes, and attributable mortality of NV-HAP using an electronic clinical surveillance definition.
Measuring the true incidence of NV-HAP is inherently challenging.Pneumonia is a syndrome that lacks a reference standard for diagnosis.3][34] Respiratory cultures are obtained in less than a third of patients treated for NV-HAP despite guideline recommendations. 35,36[42][43][44] Discharge diagnosis codes also do not provide reliable estimates of NV-HAP incidence and outcomes because they too are neither sensitive nor specific. 10,11Even cases identified using formal CDC criteria are inconsistently confirmed on external review. 7The electronic clinical criteria we applied in this study had a reasonable positive predictive value for clinically diagnosed pneumonia (81%) and moderate positive predictive value for CDC-NHSN criteria (67%).There is no reason to presume that the electronic surveillance definition is any more accurate than existing definitions since it relies on the same imperfect signs that clinicians and surveyors use to diagnose or track pneumonia.The advantage of electronic clinical criteria for NV-HAP instead is that they enable efficient surveillance in large numbers of diverse hospitals using consistent and reproducible criteria.
Most instances of NV-HAP affected clinically vulnerable patients.The median age of patients with NV-HAP was 69, most patients had multiple serious comorbidities, and one-fourth of NV-HAP events involved patients in intensive care units.The high-risk profile of the population who developed NV-HAP begs the question of the extent to which NV-HAP influenced these patients' clinical trajectories: were these patients already at high risk of death during hospitalization even without NV-HAP, or did NV-HAP lead to otherwise avoidable adverse outcomes?We incorporated a rich array of clinical parameters into our analysis to account for confounding by patients' baseline status and severity of illness using both fixed and time-varying parameters, including daily vital signs and laboratory measures.However, even after accounting for these factors, outcomes experienced by patients with NV-HAP remained worse than those without NV-HAP.
The high incidence and mortality rate associated with NV-HAP suggests it is an important hospital complication that warrants the development and testing of prevention programs.While there has been substantial work to date on defining best practices to prevent ventilator-associated pneumonia, there is very little consensus on how best to prevent NV-HAP. 457][48][49] Nurse-initiated oral care programs have suggested promising impacts on NV-HAP. 50,51However, these initiatives used manual medical record reviews or diagnosis codes to identify NV-HAP, a potential source of bias because clinical criteria are subjective and diagnosis codes are used variably by different clinicians and hospitals. 10,11Automated analyses of electronic clinical data may provide more consistent and efficient means to measure NV-HAP incidence and to track the impact of prevention programs.More broadly, our work provides a glimpse of how EHR-based surveillance has the potential to increase the number, breadth, granularity, consistency, and efficiency of quality and safety surveillance both within and between hospital systems. 52,53

Limitations
This study had limitations.There was imperfect concordance between NV-HAP surveillance criteria and confirmed pneumonia (a problem common to all clinical diagnostic and surveillance strategies due to the lack of a reference standard for pneumonia).Measurement of NV-HAP using the surveillance definition, although less subjective than diagnostic codes, could be influenced by the frequency with which vital signs and laboratory assays were measured, oxygen supplementation devices documented, variation in antibiotic prescribing and chest imaging thresholds between different clinicians and hospitals.We did not incorporate radiographic results into our surveillance definition, which may limit the definition's specificity.Developments in natural language processing and image processing may make this more feasible and could improve the measure's accuracy in the future. 54,55While our analysis of attributable mortality applied state-of-the-art analysis to detailed clinical data from 2 large national hospital networks, as in any observational analysis, there may remain unmeasured confounding, including changes in patient populations or care processes over time, particularly for the 7% of the population that coincided with the SARS-CoV-2 pandemic.We currently do not know what proportion of NV-HAP is truly preventable and hence the potential benefit of NV-HAP prevention programs.It will require prospective, real-world intervention studies, preferably randomized and multi-center, to determine the true extent to which NV-HAP can be prevented, the impact of doing so on mortality for both susceptible patients and hospital populations, and whether NV-HAP should be viewed as a measure of safety or quality.However, our study represents an early step forward by offering an estimate of incidence across diverse hospitals, associated mortality, and a generalizable surveillance method for NV-HAP.

Conclusions
In a national study involving 2 large hospital networks, we report a robust estimate of the incidence of NV-HAP in US hospitals and its possible contribution to hospital-wide mortality.Our findings underscore the importance of developing and validating robust measurement tools to monitor NV-HAP incidence and the need to identify effective prevention strategies and assess the impact of prevention initiatives on patient and hospital-level outcomes.

Figure 2 .
Figure 2. Incidence of NV-HAP Events per 100 Hospital Admissions for Each Facility280

Figure 3 .
Figure 3. Incidence of Inpatient Death Up to 60 Days Follow-Up Under Hypothetical Elimination of NV-HAP vs Current Care

1 . 4 .
Data Definitions Statistical Code eAppendix 2. Medical Record Review Guide eAppendix 3. Detailed Description of Weighted Analysis eAppendix Results of Medical record Review and Secondary Analyses SUPPLEMENT 2. Data Sharing Statement 14n-ventilator-Associated Hospital-Acquired Pneumonia in US Hospitals The VA network is the largest integrated health care network in the US and includes hospitals in all 50 US states.All VA hospitals use the same clinical EHR, Veterans Health Information Systems and Technology Architecture (VistA).Data were accessed from VistA through the Veterans Informatics and Computing Infrastructure, a computing environment that stores clinical data for research purposes.14HCAHealthcare is the largest private hospital network in the US and includes large, medium, and small hospitals in 21 US states.The majority of HCA Healthcare hospitals use a common EHR with data that are aggregated and validated centrally.HCA's centralized data has JAMA Network Open.2023;6(5):e2314185.doi:10.1001/jamanetworkopen.2023.14185(Reprinted) May 18, 2023 2/14 Downloaded From: https://jamanetwork.com/ on 09/29/2023 networks.

Table 1 .
Characteristics of All Hospitalized Patients and Patients With NV-HAP a b Categories included in the other classification are as follows-for HCA: American Indian or Alaskan Native, Hawaiian or Pacific Islander, Hispanic, Multiracial, other, unknown; for VA: American Indian or Alaska Native, Native Hawaiian or other Pacific Islander, mixed, or missing.

Table 2 .
Clinical Characteristics and Outcomes of Patients With NV-HAP