Relationship between Annualized Case Volume and Mortality in Sepsis: A Dose–Response Meta-analysis

IN the past few decades, a number of studies have evaluated the relationship between hospital volume and patient outcomes in varieties of medical and surgical conditions.^1–4  Several studies have also evaluated the volume–outcome relationship in critically ill patients.^5–7  Sepsis is a leading cause of death among critically ill patients in the United States, which puts an enormous strain on the healthcare system and is becoming a pressing public health crisis.⁸  In the past decades, the incidence of sepsis has rapidly increased worldwide,^9–12  while the mortality from sepsis has steadily decreased over time despite the absence of novel therapeutics.^13,14  Early recognition, treatment, and advances in processes of care mainly contribute to the declining mortality.¹⁵  The number of septic patients treated in a hospital (i.e., annualized case volume or hospital volume), reflecting experience, may also partly be associated with this decline. However, the relationship between annualized case volume and mortality in patients with sepsis has not been fully understood.

Previous studies have evaluated the relationship between annualized case volume and mortality in patients with sepsis but have shown inconsistent results.^16–25  Therefore, we performed this meta-analysis with the following objectives: (1) to examine the relationship between annualized case volume and mortality in patients with sepsis, hypothesizing that higher annualized case volume may lead to lower mortality; and (2) to investigate the dose–response relationship between annualized case volume and mortality in sepsis.

This meta-analysis was conducted and reported according to the Meta-analysis of Observational Studies in Epidemiology checklist.²⁶ 

We searched PubMed and Embase from inception to July 1, 2015, without any restrictions. Search terms included those related to volume, sepsis, and their variants. Details of the search strategy are available in appendix 1. We also manually searched reference lists of all the included studies as well as relevant review articles. Additionally, we also reviewed conference abstracts for unpublished work but failed to identify any eligible studies for inclusion.

Two authors (W.-J.G. and X.-D.W.) independently undertook the searches. After the removal of duplicates, we screened the titles and abstracts for relevance and accessed full text to identify eligibility of studies for inclusion in this meta-analysis. We included any observational studies (such as cohort study, case–control study, or cross-sectional study) that examined the relationship between annualized case volume and mortality among adults with sepsis, severe sepsis, or septic shock in the intensive care unit (ICU), emergency department, or hospital.

Data extraction was performed by W.-J.G. and was confirmed independently by two other authors (J.Z. and F.W.). The following information was extracted from each included study: first author, year of publication, country where the study was done, sepsis identification, clinical setting, number of sepsis cases, overall mortality rate, volume grouping (i.e., tertiles, quartiles, quintiles, or other groupings), volume category (i.e., the category according to the various cutoff values of annualized case volume), multivariate adjusted risk estimates for each category, study design, and covariates in the fully adjusted model. Extracted data were entered into a predefined standardized Excel (Microsoft Corporation, USA) file. The supplementary files of included studies were also examined for data extraction.

Quality assessment was evaluated by using the Newcastle–Ottawa Scale,²⁷  which is a validated scale for assessing the quality of nonrandomized studies in meta-analyses. This scale awards a maximum of nine stars to each study: four stars for selection of participants and measurement of exposure, two stars for comparability, and three stars for assessment of outcomes and adequacy of follow-up. We assigned scores of 0 to 3, 4 to 6, and 7 to 9 for low, moderate, and high quality of studies, respectively.

Multivariate odds ratios (ORs) and corresponding 95% CIs between extreme levels of annualized case volume (highest vs. lowest) were pooled using a random-effects model, accounting for clinical heterogeneity. Heterogeneity across studies was assessed by using the Q statistic with its P value and I² statistic.^28,29  The I² statistic is used to quantify the proportion of total variation in the effect estimation that is due to between-study variation. An I² value greater than 50% indicates significant heterogeneity. Distribution between annualized case volume and mortality in each category was described using a scatter plot. The analysis requires data on the median or mean level of annualized case volume in each category (X-axis) and the corresponding mortality rate of that category (Y-axis). After excluding two studies because of no available data,^18,20  8 of 10 studies were included for the analysis.^{16,17,19,21–25}  Publication bias was assessed by visually inspecting a funnel plot and also evaluated by using the Egger test.³⁰ 

To explore the possible sources of heterogeneity among studies, we carried out the following subgroup analyses: study design (cohort vs. cross-sectional), sepsis identification (administrative coding vs. clinical screening), quality of the risk adjustment (administrative risk adjustment vs. case-mix adjustment),³¹  clinical setting (ICU vs. emergency department vs. hospital), annualized case volume grouping (quintiles vs. quartiles vs. tertiles), sepsis with malignancies (yes vs. no), adjusted hospital region and/or teaching status (yes vs. no), sample size (greater than 50,000 vs. less than 50,000), and study quality (high vs. moderate). To test the robustness of the relationship, we also conducted a sensitivity analysis, which was investigated by omitting one study at each turn and examining the influence of each individual study on the overall risk estimate (the “leave one out” approach). A two-sided P value less than 0.05 was considered statistically significant. All analyses were performed using Stata statistical software version 13.0 (StataCorp, USA).

We used the methods proposed by Greenland and Longnecker³²  and Orsini et al.³³  for the dose–response analysis. We computed study-specific ORs and 95% CIs from the natural logarithms of the ORs and CIs across categories of annualized case volume. The method requires that the amount of annualized case volume, distributions of cases and noncases, and ORs with 95% CIs are known for at least three quantitative categories of annualized case volume. We excluded three studies from trend estimation because of less than three quantitative categories¹⁶  or no available data.^18,24  The median or mean level of annualized case volume in each category was assigned to the corresponding OR for each study. If the median or mean level was not provided and reported in ranges, we estimated the midpoint in each category by calculating the average of the lower and upper boundaries of that category. If the highest category was open ended, the midpoint of the category was set at 1.5 times the lower boundary of that category. If the lowest category was open ended, the midpoint of the category was set at 0.5 times the upper boundary of that category. All analyses were performed using R software (http://www.R-project.org; accessed April 5, 2016).

Figure 1 shows the study selection process. Our initial search yielded 148 records. After exclusion of duplicates and screening the titles and abstracts, we obtained 16 full-text articles. After reviewing the full text, 10 studies met the inclusion criteria and were included in the qualitative synthesis.^16–25  We included nine studies in the quantitative meta-analysis.^16–23,25  One study that did not report available data for analysis was excluded from the quantitative meta-analysis.²⁴ 

The characteristics of the included studies are shown in table 1. Eight of 10 were retrospective cohort studies,^16,19–25  and 2 of 10 were cross-sectional studies.^17,18  These studies were published between 2007 and 2015. The number of participants in the studies ranged from 3,437 to 1,213,219. The meta-analysis consisted of 3,495,921 participants and 834,009 deaths. Grouping and category of annualized case volume showed wide variability, but adjusted estimates could be determined for all studies even though the adjusted factors were slightly different. Details of the quality assessment of included studies are outlined in table 2. The average score was 6.7, and the score for each study was 5 or above, suggesting that all the studies were of moderate or high quality.

Nine studies that provided available data were included for quantitative meta-analysis.^16–23,25  Annualized case volume was inversely associated with mortality; that is, higher annualized case volume was associated with lower mortality (OR, 0.76; 95% CI, 0.65 to 0.89; P = 0.001; fig. 2), with significant heterogeneity (I² = 96.6%). Figure 3 shows a scatter plot describing the distribution between annualized case volume according to the median or mean level of annualized case volume in each category from one of the eight studies and the corresponding mortality rate of that category.

To explore the potential source of heterogeneity across studies, we carried out several subgroup analyses, as shown in table 3. Variations in cutoff values of category for annualized case volume across studies may be the main reason for the overall heterogeneity. The relationship between annualized case volume and mortality was consistent in all subgroups except for sepsis identification by clinical screening, unadjusted hospital region and/or teaching status, and moderate study quality. To further confirm the robustness of the results, we conducted a sensitivity analysis. Exclusion of any individual study from our meta-analysis in figure 2 did not meaningfully change the magnitude or direction of the summary effect for the relationship between annualized case volume and mortality (appendix 2).

Seven studies were included in the dose–response analysis.^16–21,23  We observed a nonlinear dose–response relationship between annualized case volume and mortality (P for nonlinearity less than 0.001). Figure 4 shows the effect of annualized case volume as a continuous variable on mortality in patients with sepsis. Reduction in mortality is seen throughout the distribution, as annualized case volume increases from the lower extreme to the higher extreme; however, the mortality benefit of a higher annualized case volume tapers off around 400 cases, which is clearly evident in figure 4.

Although the Egger test suggests that there was no evidence of publication bias (P = 0.08), it is hard to rule out the existence of publication bias by visual inspection of the funnel plot (fig. 5) since only nine studies were included in the quantitative meta-analysis.

To the best of our knowledge, this meta-analysis is a more comprehensive update that systematically and quantitatively evaluates the relationship between annualized case volume and mortality in patients with sepsis. This meta-analysis supports the hypothesis that higher annualized case volume is associated with a lower mortality. This relationship seems to be robust and consistent in subgroup analyses. Furthermore, we identified a nonlinear dose–response relationship between annualized case volume and mortality in sepsis.

Two previous reviews investigated the relationship between ICU volume and outcome and found an inverse volume–mortality relationship in critically ill adult patients.^34,35  The latest systematic review published by Nguyen et al.³⁶  was more recent and comprehensive than the two aforementioned meta-analyses.^34,35  However, the Nguyen systematic review recruited heterogeneous patient groups (e.g., cardiovascular, respiratory, hepatogastrointestinal, sepsis, etc.) rather than just all critically ill patients and the other two included critically ill patients without highly selective conditions, which limited their applicability to patients with sepsis and made their findings even more difficult to put into perspective. To better specify the volume–mortality relationship in selected patients with sepsis, we performed a meta-analysis to assess the effect of annualized case volume on mortality in sepsis. Differences between our meta-analysis and the previous reviews should be noted. First, our meta-analysis paid attention to only patients with sepsis, which was highly homogeneous and selective. Second, we also unpacked the heterogeneity across studies by examining key subgroups based on study design and patient characteristics. Last, we found a nonlinear dose–response relationship between annualized case volume and mortality, which had not been investigated in previous reviews and may strengthen causal inference.

Although the volume–outcome relationship was first reported in 1979,³⁷  the underlying mechanism has not been fully understood and requires further exploration. There are three potential mechanisms of the volume–outcome relationship.^38–40  First is “practice makes perfect” (the increased frequency of encounters allows higher case volume centers to develop more experience and streamline processes to improve quality of care), which means that high volume leads to high quality. Second is “selective referral pattern” (patients disproportionately seek care at, and physicians refer to, hospitals known for high quality of care), which means that high quality leads to high volume. Third is “organizational structural factors” (high-volume hospitals possess other factors associated with improved outcome, such as high-intensity physician staffing, multidisciplinary care teams, and protocol use), which means that the volume–outcome relationship may not exist if we could properly control for these factors. In high-acuity emergent conditions, such as sepsis, the most likely mechanisms would be “practice makes perfect” and “organizational structural factors,” as septic patients are acutely and severely ill and often do not have the luxury of choosing where they are treated.¹⁷  Thus, “selective referral pattern” could play only a small role for better quality of care.

Sepsis, as a critically ill condition that heavily relies on quality of care, would certainly benefit more from high annualized case volume hospitals with sufficient supply of medical resources and experienced experts and nurses who deliver a high quality of care than from low case volume hospitals. Our meta-analysis demonstrated that there was an inverse and a nonlinear dose–response relationship between annualized case volume and mortality. These findings suggest that centralization of care units might help in patients with sepsis. However, fully implementing centralization of septic patients is impracticable and often has a high risk–benefit ratio, especially in the regional hospitals or rural areas. In addition, the existence of a volume–mortality relationship in patients with sepsis has prompted health systems to consider the regionalization of critical care through the creation of a tiered system. However, just as a coin has two sides, there are both advantages and disadvantages regarding regionalization. On the one hand, regionalization will improve patient outcomes and increase efficiency of care delivery⁴¹ ; on the other hand, regionalization is not without risks, and little is known about the cost and feasibility of regionalization. One recent study identified multiple barriers to the acceptance and implementation of regionalization strategy.⁴²  Thus, health systems should weigh the trade-off between the potential benefits and possible risks associated with regionalization when considering the strategy of regionalized critical care.

Considering the current controversies and challenges, more studies on the volume–outcome relationship in patients with sepsis are still needed. First, further studies are required to uncover other potential mechanisms beyond “practice makes perfect” and “selective referral pattern,” such as “organizational structural factors,” including higher nurse-to-patient ratios, multidisciplinary care models, and protocols. Next, future guidelines are needed to standardize the cutoff values for annualized case volume in studies on the volume–outcome relationship in sepsis, which could overcome considerable heterogeneity and allow for assimilation of data, and then translate the evidence into clinical practice and policy implications. Finally, although there is a nonlinear dose–response relationship between annualized case volume and mortality in sepsis, the optimal volume threshold is essential for better outcomes. Two studies have shown worsening outcomes once the volume rose above a certain threshold.^43,44  One study suggested that mortality did not decrease further after a certain volume (n = 450 cases) was reached as volume increased.⁴³  This phenomenon was also observed in another study that found that after a volume threshold (n = 711 cases), there was no further mortality benefit.⁴⁴  Thus, it is important to maintain an optimal volume matched with the available resources.

Our meta-analysis has several strengths. We conducted this meta-analysis by exhaustive search without any restrictions and reported it following the Meta-analysis of Observational Studies in Epidemiology guidelines. We used the estimates from the most fully adjusted models for each study to reduce the potential of confounding. We performed several subgroup analyses to explore the potential sources of heterogeneity and evaluate the robustness of the relationship. In addition, we investigated a nonlinear dose–response relationship between annualized case volume and mortality, which can help to quantify the relationship and to test the shape of this possible relationship.

Our meta-analysis has several caveats that affect the interpretation of the results. First, we found considerable heterogeneity across the studies in our meta-analysis. It was not surprising given the differences in the data source, study population, annualized case volume grouping and category, and study design. These factors could affect our results. Nevertheless, we used a random-effects model to pool multivariate estimates, which could reduce the bias to some extent. Furthermore, dose–response analysis supports our results. Second, the results still could have been biased since most studies were retrospective and therefore limited the ability to control for confounding. However, most studies in our meta-analysis fully adjusted the important confounders. Last, although we conducted a comprehensive search of literature, it is hard to rule out the existence of publication bias since only nine studies were identified, and current guidelines do not recommend testing for funnel plot asymmetry in analyses of fewer than 10 studies.⁴⁵ 

In summary, the present meta-analysis confirmed our hypothesis and provided strong evidence for an inverse and a nonlinear dose–response relationship between annualized case volume and mortality in patients with sepsis. Variations in cutoff values of category for annualized case volume across studies may be the main reason for the overall heterogeneity. To clarify the volume–mortality relationship in patients with sepsis, future studies should uncover the mechanism of the volume–mortality relationship and standardize the cutoff values of category for annualized case volume.

Support was provided solely from institutional and/or departmental sources.

The authors declare no competing interests.