Patients and Methods
We analyzed data collected in the French CUB-REA (Collège des Utilisateurs de Donnés en Réanimation) database between 1998 and 2010. Since 1992, the CUB-REA network has collected prospective data from 41 ICUs in the urban area of Paris using a standardized web-based case report form. AECOPD were identified over the 13-year period (1998–2010) using the International Statistical Classification of Diseases and Related Health Problems. Recorded variables and coding methods were updated during annual meetings, and quality controls were performed regularly. The database quality has been assessed in previous studies. The CUB-REA project was approved by the Comité National Informatique et Liberté.
Selection of ICUs and Patients
From the 35 units continuously participating in the CUB-REA database during the study period, four mainly surgical units were excluded. Patients were identified as being admitted with AECOPD if they fulfilled the following criteria: medical (i.e., nonsurgical) admission, age > 40 years, ARF, and COPD ICD-10 codes (J96.0 + at least one of the following: J961+0; J448; J449; J441; J440; J410). For patients with multiple ICU admissions during one hospital stay, only the first was analyzed.
Recorded Data
For each patient, the following data were extracted: age, gender, Simplified Acute Physiology Score (SAPS II), associated diagnoses, organ failure support during the ICU stay (need for vasopressor support or renal replacement therapy), outcomes (ICU and hospital length of stay), method of ventilatory support (InV or NIV), and duration of MV. Mortality and standardized mortality ratio (SMR), which is used to compare mortality of patients from different hospitals, as well as the Charlson score, an applicable method for classifying comorbid conditions, were also extracted.
Case-volume Calculation
Overall volumes of admissions and of AECOPD were collected for each participating unit. Since annual admission volumes of patients with AECOPD might vary from 1 year to the next, we used a running mean according to the following formula: (volumeyear n–1 + volumeyear n)/2. Running annual volumes were then categorized into tertiles for analysis. Thus, depending on the considered year of admission, an ICU was not always classified within the same tertile. The annual prevalence of critically ill patients with AECOPD-related ARF was computed by dividing the crude AECOPD-related ARF volume by the corresponding number of total admissions.
Statistical Analysis
Continuous variables were expressed as mean ± SD and categorical variables as number and percentage. Trends over time were built for the main variables of interest, i.e., age, gender, SAPS II, MV and its type (InV alone, NIV alone, and InV + NIV), ICU and hospital length of stay, and mortality. These main variables were then analyzed by chi-square test for trend (categorical variables) and Spearman's correlation coefficient test (continuous variables). The volume-outcome relationship was explored using funnel plots where targets were the overall ICU mortality and NIV rates in the whole dataset. The case-volumes were divided into tertiles for subsequent regression analyses. Factors associated with ICU mortality and NIV use were investigated by hierarchical logistic regression (mixed model), which allows the simultaneous assessment of patient-related risk factors and ICU characteristics, while controlling for center effects. Multivariate logistic regression analyses were undertaken, with outcomes (death or NIV) as dependent variables and case-volume and all factors associated with outcomes at a p < 0.20 significance level in univariate analysis as candidate explanatory variables.
In order to account for differences in patient characteristics between case-volume tertiles, a propensity score for being admitted to a high-volume unit was derived from a cumulative logic regression model with volume category as the ordinal dependent variable and patient characteristics on ICU admission (i.e., age, gender, type of ICU admission, admission category, organ failures, and SAPS II score) as covariates. Patients from the lower volume tertile (index stays) were matched with nearest controls issued from the upper volume tertiles if they were within two tenths of the standard error of the estimated propensity score. Annual trend was taken into account by matching index and control subjects on the year of admission. We performed matching with replacement of controls for the purpose of bias reduction. This minimizes the propensity score difference between the matched index and control subjects. The number of controls was set in accordance with the observed proportions of index and control stays. The final multivariate analyses included propensity score-matched patients in a conditional logistic regression model, with ICU mortality in AECOPD patients and NIV use in ventilated AECOPD patients as variables of interest. We assessed the sensitivity of our findings by matching quartiles instead of tertiles, leaving out each ICU admission or each year and, finally, by stratifying the multivariate analysis on deciles of the propensity score.