The study was designed as a multicenter retrospective cohort study across seven acute general hospitals in Japan: Awa Regional Medical Center, Hyogo Prefectural Amagasaki General Medical Center, Ichinomiyanishi Hospital, Kameda Medical Center, Kobe City Medical Center General Hospital, Saiseikai Yokohamashi Tobu Hospital, and Tokyo National Hospital. This study was conducted in accordance with the Declaration of Helsinki.11 Additionally, this study was approved by the Institutional Review Board (IRB) of each hospital (registration number: 2020008), and the need for written informed consent was waived by the IRB of Ichinomiyanichi Hospital. This article was prepared in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement (e-Table 1).12
The patient enrollment period depended on the storage terms of the electronic medical records in each hospital during the study period of April 1, 2008 to July 31, 2020. Inclusion criteria were age ≥40 years and hospital admittance due to COPD exacerbation ≥2 times during the study period. Hospital admittance due to COPD exacerbation was determined by an admission-precipitating diagnosis of COPD exacerbation based on the 10th revision of the International Statistical Classification of Diseases and Related Health Problems (ICD-10) code (ICD-10 code: J44.1). The exclusion criteria were as follows: empirical use of anti-pseudomonal antibiotics for multidrug-resistant PA that are available in Japan (i.e. colistin, polymyxin B, and fosfomycin).
A validation study of the patient selection based on the ICD-10 code (J44.1) was conducted at Ichinomiyanishi Hospital, Kameda Medical Center, Saiseikai Yokohamashi Tobu Hospital, and Tokyo National Hospital during the study period. Patients who were selected according to the above inclusion criteria were reviewed using the clinical charts from a respiratory physician to confirm the diagnosis of COPD exacerbation.
The following patient data were collected from the Diagnosis Procedure Combination database: age, admission date, discharge date, sex, height, weight, activities of daily living, comorbidities, tracheal intubation, and prognosis. Other patient data were collected through the review of electrical medical records: baseline COPD stage, type of inhaler used (inhaled corticosteroid, long-acting beta2-agonist, long-acting muscarinic antagonist), home oxygen therapy use, activities of daily living as assessed by the Barthel index, prior PA isolation, vital signs (systolic blood pressure, respiratory rate, heart rate), mental status, oxygen use on admission, systemic steroid therapy, and antibiotic therapy.
The primary outcome was the length of the hospital stay. The outcome was measured repeatedly for each patient, and observations were nested within individuals and hospitals. Thus, patients and hospitals were considered as clusters.
Treatment of interest
The treatment of interest comprised the empirical use of anti-pseudomonal antibiotics on admission or the next day, regardless of the dose and route of administration. Anti-pseudomonal antibiotics comprised drugs that are available in Japan, as follows: ceftazidime, cefozopran, cefepime, carbapenem (biapenem, doripenem, imipenem, and meropenem), piperacillin, piperacillin/tazobactam, aminoglycoside (gentamicin, tobramycin, and amikacin), quinolone (ciprofloxacin, levofloxacin, garenoxacin, and gatifloxacin), and aztreonam.
Based on the previous literature, the following factors were selected as potential confounding factors: age, body mass index, home oxygen therapy use (binary data), respiratory rate, heart rate (≥109 beats/min or not), oxygen use on admission (binary data), mental status (altered mental status or not), systemic steroid use on admission or the next day regardless of the dose (binary data), activities of daily living (Barthel index), and number of recurrences.4,13–18
Patient characteristics are summarized as numbers and percentages for categorical variables and as the median and interquartile range (IQR) for continuous variables. A log-linked Gamma model was used to evaluate the association between each variable and the length of the hospital stay.19 Since repeated observations were obtained from a single patient, we used the generalized estimating equation method with an exchangeable correlation structure and robust standard error estimation. Hospital-specific effects were implemented as fixed effects.
Missing data were imputed using multiple imputations by chained equations on the assumption that data were missing at random.20 Covariates of the outcome analysis (number of recurrences, age, home oxygen therapy use, oxygen use on admission, heart rate, respiratory rate, mental status, systemic steroid use, admitting hospital), the treatment variable (anti-pseudomonal antibiotic use), and the outcome variable (length of the hospital stay), in addition to systolic blood pressure as an auxiliary variable, were used to estimate the missing data.21,22 The results for each imputed dataset were aggregated using Rubin’s rule.23
We also constructed a Bayesian model in which both patient- and hospital-specific effects were implemented as random effects using the complete case dataset.24 This model can be described as follows:
where Yijk is the length of the i-th stay of patient j in hospital k, and the mean of the Gamma distribution, log(μ), is determined by explanatory variables, X, and the patient- and hospital-specific effects with a zero-sum constraint, bj and bk. We set four separate sampling chains, each consisting of 5000 samples (including 4000 samples discarded for convergence). We evaluated the sampling convergence using the Gelman-Rubin statistic (R-hat) and visually inspected the trace plot.
All statistical analyses were performed using R software version 4.0.2 (R Foundation for Statistical Computing, Vienna, Austria). For Bayesian model analyses, we used the probabilistic programming language, Stan (Stan Development Team). The script is available on GitHub (https://github.com/AkihiroShiroshita/Recurrent-COPD-AE).25