Selective Referral to High-Volume Hospitals:  Estimating Potentially Avoidable Deaths FREE

In the last 3 decades, many studies have shown that, for certain procedures and diagnoses, patients have lower mortality rates at high-volume hospitals (HVHs) than at low-volume hospitals (LVHs). Although early studies lacked sufficient case-mix adjustment, more recently, the creation of specialized databases has allowed more sophisticated—though likely still imperfect—case-mix adjustment.¹ Studies using such data also show that HVHs have lower mortality rates for some conditions.¹ With rare exceptions,² health plans and purchasers have not attempted to selectively refer patients to hospitals with low case-mix–adjusted mortality or high volume. The absence of initiatives based on actual, case-mix–adjusted hospital outcomes may reflect the limitations of hospital discharge databases in most states or evidence that random events can have as much influence on the observed mortality at an individual hospital as quality of care for some conditions.^3- 4

On the other hand, condition-specific hospital volume data can be obtained in almost all states. In addition, studies that measure mortality for groups of hospitals categorized by volume will be less influenced by random events than assessments of a single hospital.

The purpose of the current study was to identify procedures and diagnoses for which there is good evidence that a volume-outcome relationship exists and to estimate the annual number of deaths in California LVHs that could be attributed to their low volume. In addition, we determined the percentage of LVH patients who were admitted through the emergency department and the additional distance LVH patients would have had to travel to reach an HVH to evaluate clinical and practical barriers to referral to HVHs.

To identify conditions for which there is evidence of a volume-outcome relationship, we searched MEDLINE, Current Contents, and FirstSearch Social Abstracts for all articles published from January 1, 1983, to December 31, 1998. Key words were: hospital, outcome, mortality, volume, risk, and quality. All articles that reported on the relationship between hospital volume and mortality were retrieved. The references of these articles were searched for other relevant studies. Among the identified articles, we included in our analysis those that (1) used data from a period within 10 years of the current study (ie, data from 1988 or later), and (2) included more than 2 HVHs.

Because the definition of the high volume category and methodological aspects differed, we could not perform a meta-analysis.⁵ Therefore, the articles identified were grouped by the procedure or diagnosis studied, and several predetermined criteria were applied to choose the single best article for each condition. Studies that used outcome variables other than in-hospital mortality (eg, 5-year survival) and studies based on patient identification variables not available to us from the California discharge database (eg, "accident occurred within city limits" in studies of trauma) were excluded, because these features would make estimation of preventable deaths using the California database impossible.

The remaining articles for each condition were evaluated to identify the study most likely to yield an unbiased estimate of the effect of volume on mortality. This evaluation included consideration of study sample size, range of volume among the hospitals included in the study, case-mix adjustment, location of study, and timeliness of data. A scoring system was developed to frame these criteria in objective terms. Studies were scored from 0 to 5 for scientific quality based on case-mix adjustment (2 points for use of clinical variables, 1 for case-mix adjustment using age and/or sex only, and 0 for no case-mix adjustment), range of the predictor variable (2 points assigned to the study with the greatest range of the predictor variable, 1 point to the study with the second greatest range, 0 for other studies), and the number of hospitals in each volume category (1 point for studies with at least 5 hospitals in every category, otherwise 0). The studies were scored from 0 to 2 for relevance of the data based on age of the data used (1 point for data collection that ended within 5 years of the current study, otherwise 0) and country in which the study was performed (1 point for United States, because such studies would be most applicable to California; otherwise 0). The scores for case-mix adjustment and for relevance were summed to generate a final score of 0 to 7. The article with the highest score was selected for study inclusion. Before scoring the studies, it was determined that, if 2 articles tied for the highest score, total sample size would be the factor used to select the best article among those with the highest scores.

We selected a volume threshold or thresholds defining hospitals as LVHs or HVHs for each condition. For the best study for each condition, volume categories for which there was no statistically significant difference from the highest volume category were collapsed into a single HVH category. All other categories were considered LVHs. Low-volume hospitals in different volume categories with separately reported odds ratios (ORs) were not collapsed into a single category. In other words, multiple categories of LVHs with incremental ORs were permitted, and data were analyzed using category-specific ORs.

The California Office of Statewide Health Planning and Development (OSHPD) maintains a database of California hospital discharges that includes annual files with abstracts from every patient discharged in a given year from any California hospital, containing the diagnoses given to and procedures performed on the patient during each admission. Using the OSHPD database, the actual number of discharges from and deaths at LVHs in California in 1997 were determined for each condition. We calculated the number of deaths for each condition at LVHs that could be attributed to their low volume, using the data derived from the best study. The OR for death between LVHs and HVHs as derived from the best study was used to calculate the expected deaths had patients been admitted to HVHs. When LVHs were separated into multiple-volume categories, ORs were calculated and applied for each category. We subtracted the number of expected deaths from the number of observed deaths to calculate the number of deaths attributable to low volume. Ninety-five percent confidence intervals (CIs) were calculated for each OR using the SEs of the ORs.⁶ The upper and lower bounds of the 95% CI for each OR were used to calculate a 95% CI for total deaths attributable to low volume. Finally, these estimates are based on single studies and by necessity are observational. Therefore, we refer to deaths attributable to low volume as potentially avoidable deaths.

Even if purchasers and clinicians agreed on policies of selective referral, it is possible that some patients would not be clinically stable enough to be referred to HVHs. Other patients might not want to travel very far to get to an HVH. To assess the impact of these barriers to selective referral, we used the OSHPD data to determine the percentage of patients admitted to LVHs who were admitted through the emergency department. For the procedures considered, we also calculated the days from admission to the performance of the procedure for emergency admissions, on the premise that some patients who enter the hospital through the emergency department do not actually receive emergent procedures. For each patient admitted to an LVH, we also compared the distance actually traveled with the distance to the nearest HVH. Distances were determined by assuming the patient traveled from the geographic center of his or her home ZIP code to the hospital.

We identified 72 articles addressing 40 procedures and diagnoses. For 19 conditions, no studies met our inclusion criteria; for 7, studies met the inclusion criteria but also failed exclusion criteria (ie, used specialized clinical information not available from the OSHPD database); and for 14, at least 1 study met all study criteria (Table 1). Among these 14, the best study showed no relationship between volume and mortality for emergent abdominal aortic aneurysm repair, knee replacement, and acute myocardial infarction.

For the remaining 11 conditions, the best study showed a statistically significant volume-outcome relationship. These included coronary artery bypass surgery,^1,7- 16 lower extremity arterial bypass surgery,¹⁷ heart transplantation,^18- 19 pediatric cardiac surgery,^20- 21 coronary angioplasty,^22- 27 elective abdominal aortic aneurysm repair,^9,12,28- 34 carotid endarterectomy,^17,35- 42 cerebral aneurysm surgery,⁴³ esophageal cancer surgery,^44- 45 pancreatic cancer surgery,^45- 52 and overall care and treatment of human immunodeficiency virus (HIV)/acquired immunodeficiency syndrome (AIDS).^53- 55

The ORs for mortality for admission to LVHs compared with HVHs are shown in Table 2. For all conditions, the best study used case-mix adjustment data beyond demographic variables. In most cases, the additional case-mix variables used included comorbidities reported on discharge abstracts, but, in some cases, specialized clinical databases created specifically for case-mix adjustment for the condition studied were used. We also reviewed the articles available for the 7 conditions with studies that met inclusion criteria, but for which preventable death estimates could not be calculated. For 3, the best studies focused on patient subpopulations defined using specialized clinical data not available in the OSHPD database, including studies of neurotrauma (used Glasgow Coma Scale to enter patients),⁵⁷ adult intensive care unit (ICU) admissions (ICU admission is not noted in OSHPD),⁵⁹ and neonatal ICU admissions (selected patients by birth weight).⁶⁰ For 3 other conditions, subarachnoid hemorrhage,⁵⁶ hepatic cancer surgery,⁴⁵ and pelvic cancer surgery,⁴⁵ we could not use the best study because its outcome variable was 30-day mortality. For breast cancer surgery, the best study used 5-year mortality.⁵⁸ However, for all 7 conditions, higher hospital volume was associated with significantly lower mortality rates.

For the 11 conditions with significant volume-outcome relationships that can be assessed using OSHPD data, a total of 58,306 patients were admitted to LVHs in California in 1997. This represented 47.9% of the 121,609 admitted with these conditions statewide. In every case except cerebral aneurysm surgery, crude mortality rates (not adjusted for case mix) were higher at LVHs than at HVHs (data not shown).

As shown in Table 3, we estimated that 602 deaths (95% CI, 304-830) at LVHs could be attributed to their low volume. This represents 26% of deaths among LVH patients with these conditions (95% CI, 13%-37%). The conditions with the largest number of deaths attributable to low volume were coronary bypass surgery, coronary angioplasty, elective aortic abdominal aneurysm repair, cerebral aneurysm repair, and overall care and treatment of HIV/AIDS.

Table 1 shows how the studies chosen relate to other available studies for each condition. Also included in Table 1 are conditions for which deaths attributable to low volume were not calculated.^61- 77 While not all studies show a statistically significant reduction in mortality at HVHs, none of the 128 comparisons of HVHs with LVHs in Table 1 showed significantly worse mortality at HVHs. However, we cannot exclude the possibility that some negative studies were never published. Most of the studies identified were conducted at universities, which are usually associated with HVHs, and researchers at those institutions may find results that do not support a volume-outcome relationship uninteresting or implausible.

To further assess the sensitivity of our findings to the choice of best study, we also calculated the number of deaths attributable to low volume using the lowest and highest estimates in the literature for each condition, summing these across all conditions. Based on the lowest estimate for each condition, there were 513 deaths attributable to low volume vs 1042 deaths for the highest estimates.

Across all 11 conditions, 29.2% of patients admitted to LVHs were admitted emergently, but the rates of emergent admission varied by condition. Applying the condition-specific rates of emergency admission and assuming no patients emergently admitted could have been referred to HVHs lowers the estimate of potentially avoidable deaths by 31.4% (to 411). However, among LVH patients who were admitted through the emergency department, 20.6% actually traveled farther to reach the LVH than they would have had to travel to reach the nearest HVH. In addition, among patients admitted emergently to LVHs for procedures, 29.1% (3933 of 13,517) underwent their procedures more than 3 days after admission, suggesting that referral to HVHs might have been possible after admission.

Because excessive distance might preclude referral of some patients to HVHs, we determined the additional distance (if any) from the admission LVH to the nearest HVH (Figure 1). Across the state, 25.2% of patients traveled farther to the LVH they used than they would have had to go to reach the nearest HVH. Overall, 58.0% of patients could have gone to an HVH without traveling more than 16 km (10 miles) farther and 76.1% could have reached an HVH without traveling more than 40 km (25 miles) farther.

We applied the results of studies of differential mortality between HVHs and LVHs to in-hospital mortality in California and found that, based on these data, a significant number of deaths potentially could be avoided through referral of certain patients to HVHs. For 11 conditions, we estimate that 602 deaths could be avoided in California each year by moving patients from LVHs to HVHs. These figures do not include patients with 7 high-mortality conditions (neurotrauma, subarachnoid hemorrhage, adults and children needing intensive care, and breast, hepatic, and pelvic cancer surgery) for which estimates could not be obtained using OSHPD data.

We refer to these deaths as potentially avoidable, because it is not clear that selective referral for large groups of patients can be accomplished or that, even if patients can be moved, the reported mortality benefits of referral would materialize. One reason for our caution is the observational nature of the research on which our estimates are based. The reported mortality benefits from admission to HVHs vs LVHs are consistent with but not proof of a causal relationship. Furthermore, the observed results may reflect in part unmeasured differences in case mix between HVH and LVH patients, because even the best case-mix measures do not explain all variations in mortality rates.

The results of the current analysis also do not provide any information about the cause of the difference in mortality rates between LVHs and HVHs. For example, are HVHs better for some conditions because of their volume ("practice makes perfect") or does the fact that certain hospitals have better outcomes lead them to receive more referrals? The nature of the causality would be especially important to policy makers concerned about outcomes for large populations. If volume-outcome relationships reflect practical experience, then consolidation of procedures into HVHs would be expected to reduce mortality. If volume-outcome effects reflect referral to better hospitals, attempts to move large numbers of additional patients to HVHs may disrupt clinical processes or create waiting lists at HVHs. At the very least, there is no guarantee that the new patients would see the same results as previous patients. However, data from New York showed that increases in volume over time at high-quality hospitals (which occurred after public reporting of mortality rates began) did not worsen outcomes for patients who receive coronary artery bypass grafts.⁷⁸ If, on the other hand, one is concerned about outcomes for a smaller group (eg, the employees of a single company) or an individual patient, selective referral to HVHs is likely to be beneficial regardless of the reason for volume-outcome relationships. The addition of small numbers of patients to the census of an HVH is not likely to change clinical patterns, so the next few patients likely will have good clinical results.

The potential negative effects of selective referral on LVHs may lessen the overall survival benefits. For example, the loss of patients who undergo coronary bypass surgery and angioplasty at an LVH may make it impossible for the hospital to support a full-time cardiologist. Patients with myocardial infarction at that hospital would then be less likely to have timely specialist consultation when necessary. Economic considerations are important as well. In particular, designation of a small number of hospitals as preferred might limit competition (potentially resulting in higher prices and lower quality). Finally, this analysis used only California data; other studies would need to determine whether the results apply across the United States.

Because in-hospital mortality can be affected by discharge policies and the availability of suitable transition care such as nursing home beds, in-hospital mortality reductions may not translate into long-term survival benefits. Studies of 30-day mortality, which is less susceptible to differences in length of stay and discharge patterns, were available for 3 of the conditions we studied (esophageal cancer, pancreatic cancer, and carotid endarterectomy). In each of these studies, reductions in 30-day mortality at HVHs were of similar magnitude to reduction in in-hospital mortality.^35,40,45 In addition, the study of 5-year mortality in breast cancer found an OR of 1.6 for LVHs vs HVHs, suggesting that volume differences can be associated with lower longer-term mortality.⁵⁸ Furthermore, other studies found that clinical complications correlate with in-hospital mortality (eg, the need for coronary bypass surgery or the occurrence of myocardial infarction after angioplasty^24- 26). Thus, it seems likely that at least a partially reduced mortality rate would be sustained after discharge.

Any selective referral initiative would face a variety of clinical and practical barriers to implementation. Some patients need immediate treatment or are too ill to be referred to an HVH rather than a nearby LVH. However, our analyses show that only 29% of LVH patients with the 11 conditions studied are admitted emergently, and approximately 21% of these already travel farther than they would to reach an HVH. Moreover, another 29% of the patients admitted emergently for procedures had their operations more than 3 days after admission. Some of these patients probably would have been stable enough for transfer to an HVH before their procedure. On the other hand, HVHs might have too few beds to provide care to all appropriate patients.

Selective referral might also be difficult for patients and their families and might disrupt continuity of care. In addition, some patients may not wish to travel far from home.⁷⁹ While the data describing additional distances required to reach an HVH show that most patients could reach an HVH by going fewer than 16 km (10 miles) farther, some patients, particularly those living in rural areas, would have to travel substantially farther. Since most of the procedures we describe are major, some patients may be willing to accept this travel burden for the benefit of a potentially lower risk of death.

A policy of selective referral by purchasers and/or payers could be accomplished in several ways. For health plans associated with a provider network (health maintenance organizations or point-of-service or preferred provider organization plans), selective referral could be negotiated into purchaser contracts with health plans and into the plans' contracts with clinicians. For health plans using precertification of hospital admissions, promotion of selective referral could be an element of the precertification process. Medicare could use volume as a criterion for selecting hospitals for its Centers of Excellence program.²

Policy makers may wish to choose only 1 or 2 conditions for initial efforts to selectively refer patients to HVHs. The choice of conditions might vary with local priorities—balancing, for example, the potential to eliminate the large mortality rate differences observed for the small number of patients with esophageal and pancreatic cancer surgery vs the prevention of even more deaths by moving many patients needing coronary bypass surgery. Other issues, such as local supply of HVHs, would also need to be considered.

Hospital volume is not the only variable that could serve as the basis for selective referral. An obvious alternative would be to base referral on the calculation of actual condition-specific, case-mix–adjusted mortality rates for all hospitals in a state. Clinicians, employers, health plans, and government programs could then encourage referral of patients to hospitals with better outcomes. Providing public access to hospital outcome data also likely would lead to a shift in patterns of use.

Both referral based on actual outcomes and publishing outcomes are preferable to referral based on proxies for quality such as hospital volume, especially since some LVHs may have good outcomes. Unfortunately, the development of condition-specific outcome measures with adequate case-mix adjustment for all hospitals in a state has proven to be very difficult. In California, developing hospital-specific mortality rates for acute myocardial infarction alone took more than 4 years.⁸⁰ New York and Pennsylvania attempted to improve patient outcomes for coronary bypass surgery by publicly reporting hospital-specific mortality rates. This public release of information did appear to improve overall mortality.^78,81 In addition, for some procedures, the annual volume at individual hospitals will be low, further limiting the ability to obtain stable estimates of case-mix–adjusted mortality rates and increasing the possibility that some hospitals can appear to have poor results in a given year by chance alone.⁸²

Neither publication of hospital-specific results nor selective referral initiatives are common in the current health care market. It is our hope that databases from which case-mix–adjusted outcomes can be calculated will be created and will obviate the need to use less precise measures like volume. It is possible that the implementation of selective referral based on proxies for quality would provide the stimulus needed for hospitals and policy makers to support the creation of such databases.

In the meantime, our data suggest that many patients could benefit from selective referral based on the best available proxies for quality of care. Additional study of the willingness of patients to move to HVHs and of the implications for patients remaining at LVHs is needed to determine that the mortality benefits we project are achievable. However, because such initiatives have not occurred without a stimulus from payers, we believe employers, health plans, and government health care programs should actively consider policies of selective referral and call for additional research on the topic.