The effect of hospital volume on patient outcomes in severe acute pancreatitis

Data regarding the patients were obtained from the Taiwan National Health Insurance Research Database (NHIRD), which is released for research purposes by the National Health Research Institute and is one of the largest and most comprehensive databases in the world [12, 13]. Information included in the inpatient database included sex, date of birth, encrypted patient identification numbers, residential or work area, dates of admission and discharge, medical institutions providing services, the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes of diagnoses (up to five) and procedures (up to five), outcome at hospital discharge (recovered, died or transferred), order codes and hospital charges. The study was exempt from obtaining approval from the Human Subjects Institutional Review Board and informed consent from the patients due to the use of an encrypted administrative database.

AP was defined by ICD-9-CM code 577.0 in any position of the five diagnoses [12]. Severe AP (SAP) was defined primarily according to the Atlanta classification scheme [14], but was modified by the presence of intensive care unit (ICU) admission (as a surrogate of acute physiology and chronic health evaluation [APACHE] II score ≥8), organ dysfunction or failure, gastrointestinal bleeding, or local complications [12, 13]. The enrollment of the patients is shown in Figure 1. We excluded patients (n = 1,414) with biliary AP who received cholecystectomy and intensive care and had no organ failure, gastrointestinal bleeding, or local complications because these patients may have had mild biliary AP and received ICU care only after the surgery [12]. After exclusion, we enrolled 22,551 patients in the following analysis.

The annual number of SAP cases per hospital was the main exposure variable. The distribution of hospital volume versus hospital mortality per hospital-year is shown in Figure 2.

We first measured hospital volume as a continuous variable (per 1 case increase per hospital-year) to assess the effect of hospital volume on outcomes [8]. Then, the hospital volume was sorted and divided into 4 and 9 about-equal subsets, respectively, to help visualize the effect of increasing volume and for practical uses. The quartile ranges were 1, 2 − 5, 6 − 13 and ≥14 SAP cases per hospital-year, respectively. The 9-quantile ranges (corresponding proportion of hospital-year units) were 1 (22.0%), 2 (12.2%), 3 (8.2%), 4-5 (10.4%), 6-8 (11.1%), 9-12 (10.1%), 13-19 (10.7%), 20-33 (9.3%), and ≥34 (5.9%) SAP cases per hospital-year, respectively. Patients were allocated essentially in decentile except that the first two decentiles with 1 case per hospital-year were collapsed into one subset due to the skewed nature of the volume. The quartiles were used for both presentation and comparison of the results and the 9-quantiles were used primarily to show the trend of volume effect on hospital mortality.

Two levels of covariates, baseline and additional, were included. The baseline covariates were patient and hospital characteristics. The patient characteristics included age, sex, year of admission, urbanization (urban, suburban and rural area) [15], Charlson comorbidity index [16, 17], and causes (biliary, alcohol-related, both or others) and severity criteria of AP.

The hospital characteristics included hospital level (medical center [>500 beds], regional [250–500 beds] and district hospitals [20–249 beds]) [13], hospital ownership (public, private not-for-profit, or private for-profit) [18] and geographical location (northern, central, southern, and eastern Taiwan).

Additional covariates were employed to account for the process of care and included the following selected treatments: cholecystectomy and life-support measures (including total parenteral nutrition [TPN], hemodialysis, vasopressors and mechanical ventilation [MV]) [12].

Primary outcome was hospital mortality [12]. Secondary outcomes were hospital charges and hospital length of stay (LOS). The charges were adjusted to the 2009 price levels in United States Dollars (USD) [12].

Data were analyzed using the SAS software, version 9.1 (SAS Institute, Inc., Cary, NC, USA). Continuous variables are presented as median (interquartile range, IQR); discrete ones as count or percentage. A two-tailed p value of <0.05 was considered significant.

We hypothesized that hospital volume is inversely associated with hospital LOS, charges and mortality. We assumed that the relationship is linear. In the univariate analysis, we performed ANOVA test for the linearity of scaled variables and linear-by-linear association Chi-square test for categorical data. To account for clustering, the effect of hospital volume was analyzed by using regression model with generalized estimating equations methods [19], specifying an exchangeable structure of a working correlation matrix, to construct regression models. Hospital mortality was regressed with a logit link function and hospital LOS and charges were log-transformed and then regressed with a linear link function on hospital volume. Both univariable and multivariable analyses were performed to yield the crude and adjusted estimates. In the multivariable analysis, we performed two consecutive models adjusting for the baseline covariates in model 1 and for the baseline and additional covariates in model 2. We examined the volume effect by first using hospital volume as a continuous variable and then as a categorical variable, as aforementioned. The effects of hospital volume are presented as odds ratios (OR) with 95% confidence interval (CI) for hospital mortality, and as percentage changes with 95% CI, which were calculated from the exponentiated regression coefficients minus 1 [20], for hospital LOS and charges. Model performance was assessed by using R-squared and c statistics. The variance of outcomes explained by hospital volume was assessed and compared with that of other covariates by using the coefficient of determination (r ² ) for hospital LOS and charges and, by the percentage change of -2 log likelihood (-2LL) for hospital mortality. The change in -2LL (%) was calculated by dividing the difference in -2LL values between the univariable and the intercept-only models by the corresponding value of the intercept-only model. The r ² was derived from the univariable linear regression model. We examined the estimated slope coefficients and the standard errors of the mean and found no indication of collinearity.

Table 1 shows the characteristics of the SAP patients. Hospital volume was correlated with hospital level, ownership and geographic location. Low quartile hospitals tended to be at the district level, private-for-profit ownership, and located in southern Taiwan, whereas higher quartile ones tended to be at the regional or center level, public or private not-for-profit ownership and located in the north.

More than two-thirds of the patients were treated in the highest quartile hospitals, and the proportion increased over time. With increasing hospital volume, patients tended to be older and less male-predominance, to live in urban areas and to have more complex comorbidities (Charlson-Comorbidity Index ≥3) and biliary causes. Among severity criteria of AP, organ failure was the most common, but the distributional pattern of the individual severity criteria differed. As hospital volume increased, the prevalence fell in gastrointestinal bleeding, but increased in ICU admission, organ failure and local complication.

More patients received TPN, vasopressors, hemodialysis, and MV in higher volume hospitals. Cholecystectomy was rarely performed in SAP and tended to be done in higher volume hospitals.

The crude estimates showed that median hospital LOS and charges increased with hospital volume. A similar trend was observed for hospital mortality except that the mortality slightly fell in the second volume quartile (Table 1).

When hospital volume was entered into the regression models as a continuous variable (Table 2), the unadjusted effect of volume on hospital mortality was not statistically significant. Model 1 shows that hospital volume was inversely associated with hospital mortality. After controlling for additional treatment-related covariates, the volume effect on hospital mortality attenuated and became insignificant (Model 2). Volume, as a continuous variable, appeared to have no effect on hospital LOS, but was inversely associated with lower hospital charges in the unadjusted and fully adjusted models.

When hospital volume was entered into the regression models as a categorical variable (Table 3), the results of the volume effect on hospital mortality were similar to those modeled using volume as a continuous variable. Compared to the lowest quartile hospitals, SAP patients treated in higher volume hospitals had nearly 40% less risk of hospital mortality (Model 1). Volume effect became insignificant when differences in treatment among volume quartiles were controlled. When hospital volume was divided into 9 about-equal subsets, the effect of volume on hospital mortality appeared to plateau at ≥3 SAP cases per year (Figure 3). However, the results of hospital LOS and charges differed from those using volume as a continuous variable. For example, compared to the lowest quartile hospitals, patients treated in higher volume hospitals had a shorter hospital stay after controlling for the baseline covariates (Model 1). The volume effect on hospital LOS attenuated but persisted after additional adjustment of the treatment covariates (Model 2). Conversely, the volume effect on hospital charges was significant only after controlling for the baseline covariates (Model 1).

The variances in the outcomes explained by various variables are shown in Figure 4. The extent of the effect of hospital volume was greater on hospital mortality than on hospital LOS and charges. Nevertheless, the degree of the effect associated with hospital volume on various outcomes was relatively small compared to those of most patient and hospital characteristics.