Association of Practitioner Interfacility Triage Performance With Outcomes for Severely Injured Patients With Fee-for-Service Medicare Insurance

Key Points

Question  Is variation in the quality of interfacility triage associated with the facility or the treating practitioner?

Findings  In this cohort study of Medicare claims data from 124 008 severely injured patients, practitioners were an important source of variation in interfacility triage at nontrauma centers and practitioner triage practices were associated with and patient outcomes, such that the case-fatality rate increased in association with undertriage.

Meaning  The findings suggest that practitioner variation is associated with patient outcomes and thus that efforts to eliminate preventable injuries after death should include both system-level and practitioner-based solutions.

Importance  Despite evidence that treatment of severely injured patients at trauma centers is associated with reduced mortality, nearly half of all such patients are treated at nontrauma centers (undertriaged). Little is known about whether interfacility undertriage occurs because of practitioner decision-making or institutional and regional factors.

Objectives  To assess the associations between variation in triage practitioners at nontrauma centers and between practitioner-level variation and patient outcomes after injury.

Design, Setting, and Participants  This retrospective cohort study used Medicare claims data from severely injured patients presenting to nontrauma centers and the practitioners who evaluated them in the emergency department from January 1, 2010, to October 15, 2015. Data analysis was performed from January 15, 2018, to March 21, 2019.

Main Outcomes and Measures  Proportion of variation in undertriage associated with practitioners, practitioner rates of undertriage, practitioner characteristics associated with undertriage, and 30-day case-fatality rate.

Results  A total of 124 008 severely injured patients (mean [SD] age, 81 [8.4] years; 67 253 [54.2%] female) and the 25 376 practitioners (5564 [21.9%] female) who evaluated the patients in the emergency department of nontrauma centers were included in the study. Undertriage occurred among 85 403 patients (68.9%), with 40.6% of total variation associated with practitioners, 37.8% with hospitals, and 6.7% with regions. Compared with physicians with National Provider Identification (NPI) enumeration before 2007, those with an NPI enumerated between 2007 and 2010 had an undertriage risk ratio (RR) of 0.98 (95% CI, 0.97-0.99), and those with an NPI enumerated after 2010 had an undertriage RR of 0.96 (95% CI, 0.94-0.99). Hospitals with neurosurgeons had an undertriage RR of 1.51 (95% CI, 1.45-1.57) compared with those that did not; hospitals with spine surgeons had an undertriage RR of 1.10 (95% CI, 1.06-1.13); hospitals with general surgeons had an undertriage RR of 1.13 (95% CI, 1.09-1.17). Compared with practitioners who undertriaged 25% or less of patients, a statistically significant increase was found in the odds of death for patients treated by practitioners with a triage rate of less than 25% to 50% (odds ratio [OR], 1.08; 95% CI, 1.05-1.20) and less than 50% to 75% undertriage (OR, 1.12; 95% CI, 1.09-1.26) but not undertriage at greater than 75% (OR, 1.03, 95% CI, 1.00-1.18). In sensitivity analyses to adjust for unmeasured confounding, the association between triage practices and the case fatality rate became monotonic; compared with patients treated by practitioners with an undertriage rate of 25% or less, the odds of case fatality were 1.13 (95% CI, 1.05-1.21; P = .001) among patients treated by practitioners with undertriage rates less than 25% to 50%, 1.22 (95% CI, 1.13-1.32; P < .001) for patients treated by practitioners with undertriage rates less than 50% to 75%, and 1.20 (95% CI, 1.10-1.30; P < .001) for patients treated by practitioners with undertriage rates greater than 75%.

Conclusions and Relevance  The findings suggest that individual practitioner practices are an important source of variation in triage and represent a potential locus of intervention to reduce preventable deaths after injury.

In 2014, a total of 30 000 patients experienced preventable deaths after injury.¹ Triage of severely injured patients to designated trauma centers is associated with reduced morbidity and mortality among patients with traumatic injury.^2-5 Long-standing efforts to promote systematic triage of severely ill patients to trauma centers include the development of infrastructure necessary to support a tiered delivery system for trauma, accreditation standards that require trauma centers to perform outreach to nontrauma centers, and the development of educational programs to disseminate clinical practice guidelines.^6,7 Despite these efforts, approximately half of all severely injured patients receive definitive care at a nontrauma center rather than a trauma center.^8-11

Factors associated with the failure to transfer patients to trauma centers (undertriage) include those related to the patient (eg, age, preferences), hospital (eg, availability of resources), and region (eg, distribution of aeromedical transport services).^9,10,12,13 The contribution of practitioners in the emergency department (ED) of nontrauma centers to undertriage is, to our knowledge, an unexplored area that could inform quality improvement efforts. The present study used patient-level claims data from the US Medicare program to quantify variation in interfacility triage associated with practitioners and to examine the association between practitioner triage practice patterns and patient outcomes.

We performed a retrospective cohort study of patients with severe traumatic injuries and the practitioners who treated them in EDs using January 1, 2010, to October 15, 2015, claims data from the Centers for Medicare & Medicaid Services (CMS). Medicare claims serve as a useful source of information for 2 reasons. First, elderly people experience high rates of undertriage and case fatality after injury, allowing analysis of the consequences of variation in care.^4,12,14 Second, Medicare claims allow the longitudinal tracking of patients and the identification of practitioners providing care. The University of Pittsburgh Institutional Review Board approved the study protocol and granted a waiver of informed consent. Data were not deidentified.

We obtained patient-level data from the 2010-2015 Medicare Beneficiary Summary Files, Inpatient and Outpatient Standard Analytic Files, and Carrier Files. These files contain the demographic characteristics and administrative health records for all fee-for-service Medicare beneficiaries and the National Provider Identification (NPI) number of the billing practitioner. We selected claims that involved a moderate to severe traumatic injury (defined as an abbreviated injury score [AIS] ≥3) using validated International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnostic codes.^15,16 Practitioner-level data came from the CMS National Plan and Provider Enumeration System, hospital-level data from the American Hospital Association Annual Survey, and the trauma center designation of hospitals from the Trauma Information Exchange Program.

We used the Inpatient and Outpatient Standard Analytic files to identify visits to acute care, nonfederal hospitals for the treatment of severe injuries by translating AISs to Injury Severity Scores (ISSs) and selected patients with an ISS greater than 15 (Figure). We constructed episodes of care by ordering claims by day and combined visits that occurred within 1 day of each other into a single episode. Episodes began when the beneficiary presented initially and ended (1) after 24 hours or more at an acute care hospital, (2) with discharge from the ED (assuming the beneficiary did not present to another hospital within 24 hours), or (3) with death in the ED. For episodes with multiple claims from the same day, we ordered them under the assumption that patients would only move from nontrauma centers to trauma centers or from low-volume hospitals to high-volume hospitals (eAppendix in the Supplement).

We identified the treating ED practitioner using unique NPI numbers and linked NPI numbers to care episodes using 2 methods. For episodes that involved a hospitalization at the same hospital as the presenting ED, we used NPIs from the professional claims in the carrier file related to that episode (eAppendix in the Supplement). For episodes that did not involve a hospitalization at the same hospital as the presenting ED, we used the performing practitioner NPI in the Outpatient Analytic File.

We abstracted patient demographics and vital status at discharge and 30 days after presentation from the Medicare Beneficiary Summary file. Information on comorbid conditions, injury characteristics, and organ failure came from ICD-9-CM diagnosis codes in the claims. We calculated the ISSs by mapping ICD-9-CM codes to AISs.

The CMS National Plan and Provider Enumeration System provided information on practitioner characteristics, including sex, type, and experience (eAppendix in the Supplement). We also determined the hospitals at which practitioners worked and categorized practitioners as working at both level 1 and 2 and nontrauma centers or working at only nontrauma centers.

We obtained the trauma center status of hospitals from the Trauma Information Exchange Program and categorized hospitals as trauma centers (level 1 and 2 centers) or nontrauma centers (level 3-5 or nondesignated centers). The American Hospital Association Annual Survey provided information on institutional characteristics (eg, number of intensive care unit beds). We used ICD-9-CM–specific procedure codes to identify the presence of surgical and radiographic resources at these hospitals (eAppendix in the Supplement). We also summarized the proportion of ED care at each hospital that was provided by emergency medicine (EM) physicians and described hospitals as being EM staffed if more than two-thirds of care was provided by physicians with EM certification.⁹ Finally, we identified regions by their hospital referral region identifier, and we determined the number of level 1 and level 2 trauma centers in each region and categorized them as having 0, 1, or more than 1 trauma center.

We excluded episodes of care that began at level 1 and 2 trauma centers (because these hospitals had the ability to provide definitive care of injuries) and episodes that began at critical access hospitals (because they would never have the ability to provide definitive care for severe injuries). We also excluded practitioners who evaluated fewer than 10 total (moderate and severely injured) beneficiaries during the 6-year period because we could not reliably generate estimates of their performance. We categorized episodes as involving moderate (ISS <15) or severe (ISS >15 [based on data calculation method, an ISS of exactly 15 was not possible]) injuries and restricted the analyses to beneficiaries’ first episode of care each year that involved a severe injury. We defined undertriage as the failure to transfer a patient with severe injuries to a higher level of care directly from the ED or within 1 day of admission, categorizing those who died within 24 hours of arrival or who were discharged from the ED as undertriaged.

To determine the proportion of variation in undertriage associated with practitioners, we fit an empty mixed-effects logistic regression model, clustered at the region, hospital, and practitioner level, with the dependent variable of patient undertriage. We used a nested model with variance terms for each level and calculated their intraclass correlation coefficients.¹⁷ In a sensitivity analysis, we tested the influence of imperfect nesting of practitioners in hospitals by excluding patients treated by practitioners who worked at more than 1 facility.

We began by calculating nested practitioner-specific undertriage rates, defined as the proportion of patients undertriaged by a physician at each hospital at which they worked. The American College of Surgeons (ACS) recommends less than 5% undertriage as the standard for regional triage patterns.⁶ We applied a similar benchmark to the performance of hospitals and practitioners, given the high proportion of severely injured patients taken initially to nontrauma centers, but acknowledge that other standards may apply. However, because few practitioners achieved this objective, we categorized performance into 4 groups (≤25% undertriage, >25% to 50% undertriage, >50% to 75% undertriage, and >75% undertriage), an approach we selected based on the Akaike information criterion for our models. We examined practitioner-level variation graphically and compared practitioner characteristics by category of triage performance using χ² tests.

To understand practitioner factors associated with undertriage, we performed a multivariable analysis using random-effects Poisson regression. The dependent variable was a count of each practitioner's unsuccessfully triaged patients with severe injuries, with the total number of patients with severe injuries serving as the offset term (ie, adjusting for the number of exposures). We included practitioner-, hospital-, and regional-level covariates and specified a hospital-level random effect. We interpreted exponentiated coefficients in this model as risk ratios (RRs), controlling for the other covariates in the model.

To determine whether practitioners influenced outcome measures in trauma, we tested the association between practitioners’ triage practice patterns and the 30-day case-fatality rate. We fit a mixed-effects patient-level logistic regression model with the 30-day case-fatality rate as the dependent variable and the practitioners’ undertriage rate as the primary independent variable. Covariates included patient characteristics (age [by decade], sex, selected comorbidities, maximum AIS by body region, and organ failure), practitioner characteristics (sex, specialty, and years of experience), hospital resources (consultants [eg, neurosurgery], medical school affiliation, presence of residents, and EM staffing), and regional characteristics (number of trauma centers). In post hoc sensitivity analyses, we used the in-hospital case-fatality rate as the outcome variable to test whether time altered the strength of the association between clinician practices and patient outcome. We also excluded patients discharged from the ED, using this disposition as a surrogate method of additionally adjusting for patient robustness before injury. Finally, we tested the influence of imperfect nesting of practitioners (ie, adjusted for the correlation introduced by practitioners who worked at >1 hospital) to ensure that cross-classification would not lead to overestimation of the response variation in the different levels using Bayesian analyses.¹⁸

We performed data management and statistical analyses using Stata statistical software, version 12.0 (StataCorp) and MLwiN, trial version 2.36 (University of Bristol), with statistical significance set at 2-sided P < .05.^19,20 Data analysis was performed from January 15, 2018, to March 21, 2019.

We identified 1 558 072 episodes of moderate to severe traumatic injury in patients who presented initially to a nontrauma center from January 1, 2010, to October 15, 2015. We identified a treating ED practitioner for 1 159 160 patients (74.4%), representing 132 556 unique practitioners (eTable 1 in the Supplement gives characteristics of unmatched patients). We excluded patients with only moderate injuries and those treated by practitioners who evaluated fewer than 10 total patients (eTable 2 in the Supplement gives practitioner characteristics). The final analysis included 124 008 severely injured patients (mean [SD] age, 81 [8.4] years; 67 253 [54.2%] female) treated by 25 376 practitioners (5564 [21.9%] female) (who saw at least 10 beneficiaries) at 2461 acute care hospitals. Table 1 gives the characteristics of patients, practitioners, and hospitals in the final sample.

We observed broad variation in the triage of trauma patients: 38 605 (31.1%) were triaged successfully, and 85 403 (68.9%) were undertriaged (1337 died in the ED, 28 583 were discharged, and 55 483 were admitted to the hospital where they presented). On the basis of intraclass correlation coefficients, 40.6% of total variation was correlated with practitioners, 37.8% with hospitals, and 6.7% with regions. Our sensitivity analysis to account for imperfect practitioner nesting in hospitals produced similar results.

Practitioners who evaluated 10 or more total beneficiaries during the 6 years saw a median of 4 (interquartile range, 2-7) severely injured patients and undertriaged 75% (interquartile range, 45%-100%). Only 2388 practitioners (9.4%) met ACS-endorsed thresholds for successful triage. Practitioner characteristics by category of triage performance are given in eTable 3 in the Supplement.

In the multivariable clustered Poisson regression analyses, EM training was associated with reduced risk of undertriage, whereas greater experience and lower volumes of patients were associated with increased risk of undertriage (Table 2). Practitioners at hospitals with neurosurgical services (RR, 1.51; 95% CI, 1.45-1.57), spine surgeons (RR, 1.10; 95% CI, 1.06-1.13), and general surgeons (RR, 1.13; 95% CI, 1.09-1.17) were significantly more likely to undertriage patients.

After adjustment for patient, practitioner, and hospital characteristics, there was a U-shaped association between treatment by a practitioner who undertriaged more patients and odds of 30-day case fatality (Table 3). In sensitivity analysis in which patients discharged from the ED (who may have been systematically less frail) were excluded, compared with patients treated by practitioners with a triage rate of 25% or less, a statistically significant increase was found in the odds of case fatality for patients treated by practitioners with a triage rate of 26% to 50% (odds ratio [OR], 1.13; 95% CI, 1.05-1.21), 51% to 75% (OR, 1.22; 95% CI, 1.13-1.32), and greater than 75% (OR, 1.20; 95% CI, 1.10-1.30) (Table 4). The use of Bayesian analysis to adjust for imperfect nesting of practitioners within hospitals produced similar point estimates and credible intervals to the main analysis.

In a retrospective cohort study of Medicare fee-for-service patients, we found that 41% of the total variation in triage at nontrauma centers was associated with practitioners. Few practitioners met the ACS threshold of less than 5% undertriage. Practitioners who undertriaged fewer patients tended to have less experience and were less likely to work at hospitals with the resources necessary to manage injured patients definitively. In sensitivity analyses, treatment by a practitioner with low rates of undertriage was associated with a small mortality benefit.

Prior studies^21-23 have found variation in patterns of care and outcomes associated with individual practitioners in other clinical domains, including discretionary spending in the intensive care unit, length of stay for general surgical patients, and mortality after acute myocardial infarction. However, in trauma, debate still exists about the significance of clinician decision-making during triage. An argument against the role of practitioner decision-making as a mediator is that mortality is associated with many nonmodifiable variables (eg, the injury complex, the patient’s preexisting physiologic status).²⁴ The alternative argument stresses the importance of early intervention in altering the trajectory of the condition in the patient, and the institution and practitioners who first provide care are responsible for the outcome.^25-27 Our study supports the latter view. Specifically, our results suggest that the quality of the practitioners' decision-making, as manifested by their triage patterns, is associated with the case-fatality rate. Interventions that address practitioner judgment (ie, their ability to assess accurately the probability of a severe injury) and decisional thresholds (ie, their preferences for erring on the side of undertriage or overtriage in the context of diagnostic uncertainty) have the potential to reduce variation in undertriage. These results may therefore have policy implications by informing the distribution of resources intended to improve the quality of trauma regionalization, such as those authorized by legislation such as the Patient Protection and Affordable Care Act.³

The association between practitioner-level variability in triage practices and case-fatality rates may have several explanations. Patients treated by practitioners with higher triage rates are more likely to be transferred to a trauma center, where treatment is associated with a range of benefits.^2,28,29 Alternatively, triage practices may serve as a marker for other types of care that are also associated with mortality benefits (eg, early and aggressive resuscitation, appropriate use of antibiotics for patients with open fractures).^30,31 As indicated in Table 4, we also found that the association between triage practices and case fatality attenuated over time, which is consistent with a conceptual model of triage as having the greatest effect on early outcomes.

In the present study, the association between practitioner triage practices and patient outcome was U-shaped. In post hoc sensitivity analyses, this observation was shown to likely be secondary to residual unmeasured confounding. Specifically, the exclusion of patients discharged from the ED had a monotonic association between practitioner triage practices and odds of death. Patients that we identified as severely injured who were discharged from the ED may have been misclassified in our analysis. However, a study previously validated,³² through medical record review, the observation that practitioners frequently discharge severely injured patients from the ED. We consequently speculated that these discharges offered an indirect method of assessing patient robustness. Better risk adjustment allowed us to assess the association between practitioner performance and patient outcomes.

Our study has several limitations. First, although the use of Medicare fee-for-service claims allowed us to evaluate triage patterns in a national sample of practitioners, it constrained the analysis to adults older than 65 years with fee-for-service Medicare. Our observations may not be generalizable to younger patients or those with Medicare Advantage. Second, we excluded a large proportion of patients; 26% of the cohort remained unmatched—a common problem in claims-based research about ED practice.^33,34 As indicated in eTable 1 in the Supplement, the case-fatality rate was increased among these patients, and they were less likely to be transferred to a trauma center. We speculate that this exclusion would therefore bias our estimates toward the null (if unmatched patients occurred systematically among specific practitioners) or would have no effect (if unmatched patients occurred randomly). We also excluded patients treated by practitioners who treat low volumes of patients to ensure the stability of our estimates of performance. As indicated in eTable 2 in the Supplement, these practitioners were less likely to be EM board certified and more likely to work at smaller hospitals with fewer resources. Our observations may therefore have limited generalizability to this population.

Third, our analysis attributed responsibility for triage to ED physicians. We speculate that our failure to capture the informal and formal contributions of other practitioners to the decision-making process biases our results toward the null, underestimating the differences among practitioners. Fourth, we used ICD-9-CM–derived ISS to identify the cohort, which is an imperfect process. However, we have no reason to believe that this misspecification varied systematically. Fifth, our study is subject to survivorship bias. Specifically, patients who died before arrival to the initial triage hospital were excluded from the analysis; those who died shortly after arrival at the initial hospital were included but were treated as triage failures. A recent article by Bardes et al³⁵ reported that of the patients in the National Trauma Data Bank who died on the first day, 33% died in the first 4 hours. Other studies^36,37 reported estimates of 20% to 25%. We could not adjust for the inclusion of this cohort of patients; thus, bias may have inflated our point estimates. Related to this issue, our primary outcome variable (transfer within 24 hours of presentation) was subject to misspecification because we had only date stamps and not date and time stamps. However, misspecification was likely to be nondifferential among hospitals and practitioners and for injury severity.³⁸ Sixth, we generated estimates of hospital resources using claims filed for moderately and severely injured patients, possibly resulting in an underestimate of resources available at hospitals with high transfer rates. To mitigate this problem, we included not only claims from patients with severe injuries during their index hospitalization (n = 124 008) but also those from patients who presented with a moderate to severe injury from January 1, 2010, to October 15, 2015 (n = 3 330 072). In addition, we set a low threshold to distinguish between the presence or absence of resources (number of claims required was 10).

The findings suggest that variation in triage patterns persists despite best-practice quality improvement efforts by major stakeholders. Using Medicare claims, we found variation in practitioner interfacility triage practices and an association with patient outcomes. We believe that interventions designed to reduce undertriage by EM practitioners should be evaluated in future studies.

Accepted for Publication: July 29, 2019.

Corresponding Author: Deepika Mohan, MD, MPH, Department of Critical Care Medicine, University of Pittsburgh, 606D Scaife Hall, 3550 Terrace St, Pittsburgh, PA 15261 (mohand@upmc.edu).

Published Online: October 23, 2019. doi:10.1001/jamasurg.2019.3944

Author Contributions: Dr Mohan had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Concept and design: Mohan, Angus, Rosengart, Barnato, Fischhoff, Chang, Kahn.

Acquisition, analysis, or interpretation of data: Mohan, Wallace, Kerti, Barnato, Yealy, Chang, Kahn.

Drafting of the manuscript: Mohan, Angus, Rosengart, Yealy, Chang.

Critical revision of the manuscript for important intellectual content: Mohan, Wallace, Kerti, Angus, Rosengart, Barnato, Fischhoff, Chang, Kahn.

Statistical analysis: Mohan, Kerti, Angus, Rosengart, Chang.

Obtained funding: Mohan.

Administrative, technical, or material support: Mohan, Wallace, Angus.

Supervision: Angus, Rosengart, Barnato, Yealy.

Conflict of Interest Disclosures: Drs Mohan, Wallace, Angus, Barnato, and Kahn reported receiving grants from the National Institutes of Health during the conduct of the study. Dr Angus reported receiving personal fees from Ferring Pharmaceuticals Inc, Bristol-Myers Squibb, Bayer AG, and GenMark Diagnostics; receiving stock options from Alung Technologies Inc outside the submitted work; and having a patent to Selepressin compounds, compositions, and methods for treating sepsis pending and a patent to Proteomic biomarkers of sepsis in elderly patients pending. No other disclosures were reported.

Funding/Support: This study was supported by grants DP2 LM012339 (Dr Mohan) and NHLBI-K08-HL122478 (Dr Wallace) from the National Institutes of Health.

Role of the Funder/Sponsor: The funding agencies reviewed the study but played no role in its design, collection, analysis, or interpretation.

Additional Contributions: The Biostatistics and Data Management Core, Clinical Research, Investigation, and Systems Modeling of Acute Illness Center, Department of Critical Care Medicine, University of Pittsburgh, Pittsburgh, Pennsylvania, provided statistical support. No compensation was provided for the work.