Study design and data source
For this retrospective cohort study, we used the Japanese Diagnosis Procedure Combination Database . This database contains administrative claims data and clinical information. All 82 academic hospitals are obliged to provide information to this database. However, participation by community hospitals is voluntary. The database includes the following information: age, sex, diagnosis (primary diagnosis at admission, comorbidities at diagnosis, and post-admission complications) recorded by the International Classification of Diseases, 10th Revision (ICD-10) codes [32,33] as well as the following text data in Japanese, whether transferred by ambulance, medical procedures including types of surgery, daily records of drug administration and devices used, date of admission and discharge, and discharge status. The database was structured explicitly to differentiate between pre-admission comorbidities and post-admission complications. The dates of surgery, procedures performed, and drugs administered were also recorded. All clinical data for each patient were recorded at discharge by attending physicians.
This study was approved by the Institutional Review Board of the University of Tokyo. The need for informed consent was waived because this was a retrospective study that used anonymized data.
We identified patients in the database with pneumonia as the primary diagnosis upon admission, and a hospital discharge date between 1 September 2014, to 31 March 2017. We included patients with pneumonia-associated sepsis according to the following criteria: (i) a primary diagnosis of pneumonia on admission (i.e. coded as pneumonia in the database; excluding suspected cases) and (ii) underwent CRRT with the AN69ST membrane or a standard membrane within 2 days of admission. The exclusion criteria were as follows: (i) age younger than 18 years; (ii) death within 2 days of admission and (iii) administration of CRRT with AN69ST membrane and a standard membrane within 2 days of admission.
Exposure and endpoint
The exposure of interest was AN69ST-CRRT (AN69ST group) compared to standard CRRT (non-AN69ST group) within 2 days at admission. The primary outcome was in-hospital mortality; the 30-day mortality rate and the length of stay were secondary outcomes.
Hospital volume was defined as the average annual number of pneumonia patients who have underwent any CRRT within 2 days of admission. Comorbidities on admission were extracted for each component of the Charlson Comorbidity Index, using algorithms developed by Quan et al . Data were extracted from the ICD-10 codes of complications and the procedures listed in Additional File 1. Data on body weight and height were missing for some patients; thus, for these patients, body mass index values were missing.
We used propensity score methods, which have been used in several previous retrospective observational studies for comparing groups with similar characteristics without specification of the relationship between confounders and outcomes [35,36]. Similarly, we used propensity score matching  to adjust for differences in baseline characteristics and the severity of the condition on admission, between the AN69ST and non-AN69ST groups. To estimate the probability of receiving AN69ST-CRRT or other standard CRRT, a propensity score was calculated for each patient using multivariable logistic regression analysis. Baseline characteristics were included in the model as independent variables. All patients in the AN69ST group were 1-to-1 matched with patients in the non-AN69ST group, based on nearest-neighbor matching without replacement. The caliper was set at 0.2 for the standard deviation of the propensity scores. The balance between the two groups was compared using the standardized mean difference (SMD), and SMD less than 0.1 was considered to indicate negligible imbalance. The outcomes between the two groups were compared using Fisher’s exact test for in-hospital mortality and the Mann-Whitney U test for length of stay. Kaplan–Meier survival curves were plotted for the AN69ST group and the non-AN69ST group, and the log-rank test was used to compare the survival curves.
We conducted subgroup analyses on all baseline characteristics and in-hospital mortality using the Breslow-Day test for categorical variables and generalized linear models for continuous variables. A p-value of less than 0.05 was considered statistically significant. All analyses were conducted using SPSS version 22 (IBM Corp, Armonk, NY, US) and R version 3.1.3 (The R Foundation, Vienna, Austria).