Study design and data source
We conducted a multicenter observational study using EMRs of 55 hospitals in the NHO database managed by the NHO Headquarters. The NHO has been established in April 2004 and is the largest hospital organization in Japan, including general acute care hospitals and specialized long-term care hospitals. The NHO has the administrative claims database (Medical Information Analysis databank; MIA) and the clinical information database (NHO Clinical Data Archives; NCDA) . The MIA is the claims database based on the Diagnosis Procedure Combination/Per-Diem Payment System (DPC/PDPS) of case-mix patient classification and a lump-sum payment system for patients in Japan [12, 13]. The MIA contains patient information of age, sex, diagnosis, comorbidities, complications, medical procedures, medications, etc. based on the medical insurance system. The NCDA is based on the Standardized Structured Medical Record Information Exchange (SS-MIX), including medical charts, laboratory data, bacterial culture data, and many other data fields on daily basis [14, 15]. Additional details have been reported elsewhere [12, 13, 15, 16].
Patients who had a hospitalization between April 2016 and March 2020 were included. Inclusion criteria were patients aged 15 years and older, and having a diagnosis of pneumonia, urinary tract infection (UTI), biliary infection, or sepsis with the administration of the intravenous antibiotics. The diagnoses were defined according to the International Classification of Diseases, 10th Revision (ICD-10) codes: pneumonia (A241, C349, J15, J18[0–289], J690, J85, J958, J170), UTI (N1[0–2], N209, N390, T835), biliary and pancreatic infections (K8[0–7]), and sepsis (A241, A327, A415, A41, I301, I330, J209, J950, L029, M8699, T814). Bacterial culture tests positive from either specimen of urine, blood, or sputum after more than 3 days of hospitalization were defined as healthcare-associated infection. The first positive of the carbapenem-resistant pathogen from the bacterial culture test was considered the index positive result of the carbapenem-resistant (CR) bacteria. Patients who had tested positive for CR bacteria at least once were classified into the CR infection group. Patients whose culture tests were positive for only carbapenem-susceptible bacteria were classified as carbapenem-susceptible (CS) infection group. Carbapenem resistance was determined based on the results of antimicrobial susceptibility testing according to the JANIS definition .
Some patients received special medical care with the public expenditure, such as incurable diseases, congenital diseases, and war victims, whose medical costs are fully covered by the government. In this study, none of the patients treated by the public expenditure met the inclusion criteria: in the NHO database, the number of publicly funded patients with carbapenem-susceptible or -resistant organisms detected by bacterial culture tests was 0.03% (305 of 88,099 in total at the first step of the patient inclusion flowchart in Fig. 1), and there were no relevant patients in our final analysis.
We hypothesized that patients with CR infections would stay in the hospital longer and that their cost during hospitalization would be higher than that of patients with CS infections. The outcomes were total direct cost during hospitalization, in-hospital mortality, and total LOS.
We extracted baseline of patient characteristics, medications, and medical procedures from the database: age, sex, body mass index (BMI), route of hospital admission, antibiotics, immunosuppressive drugs including chemotherapy and steroids, intensive care unit (ICU) admission, surgical procedure, dialysis, mechanical ventilation. In addition, we extracted the clinical data from the database related to bacterial infections: results of culture tests with organism type and antimicrobial susceptibility testing, white blood cell (WBC) count, body temperature (BT), and C-reactive protein (CRP) levels during hospitalization. The Charlson Comorbidity Index (CCI) scores were calculated according to the Quan’s coding algorithms , and used as a measure of the burden of chronic illness . Imputation methods were used to estimate missing values in body weight, height, CRP level, WBC count, and BT. The following bacteria species were defined as research targets: Pseudomonas aeruginosa, Acinetobacter spp., Escherichia coli, Klebsiella pneumoniae, Klebsiella oxytoca, Klebsiella aerogenes, Enterobacter cloacae, Citrobacter freundii, Proteus mirabilis, and Proteus vulgaris. Following the JANIS guidelines [8, 10], the definition of resistance was based on the bacterial culture test. The carbapenem-resistance was defined by the result of antimicrobial susceptibility testing, as one of the above organisms with at least one determination of an ‘R’ (resistant) in the two algorithms (‘R’ result to imipenem-cilastatin sodium, or ‘R’ result to both meropenem hydrate and cefmetazole sodium). Hospital-level characteristics were grouped according to the number of beds: < 400, 400–500, and ≥ 500 beds.
To summarize patient characteristics, continuous variables were expressed as the mean and standard deviation (SD) or the median and interquartile range (IQR), depending on the distribution of variables. The Wilcoxon rank-sum test or the Welch test were used for assessing between-group differences. Categorical variables were expressed as proportions and compared using the chi-square test  To assess the impact of carbapenem resistance on in-hospital mortality, LOS, and cost, the hierarchical regression models were used with random effects accounting for the difference among hospitals [6, 19]. The covariates of interest were used for the adjustment in the models. The impact of carbapenem resistance on LOS and hospitalization cost was estimated as percent changes by linear regression with log-transformed values of these outcomes, and the impact on in-hospital mortality showed as odds ratios by logistic regression. The final regression model included age, sex, BMI, CCI, immunosuppressive drugs, antibiotic use before the bacterial culture test, ICU admission, undergoing a surgical procedure, disease type, and death, based on the methods in previous studies [4,5,6]. When death was the outcome of the model, it was not included as a covariate. The Akaike information criterion (AIC) was used for comparison among the unadjusted, adjusted, and multilevel models. To ensure the robustness of our results, we additionally adjusted for confounding by the inverse probability of treatment weighting (IPTW) with the method of the propensity scores overlap weights [20,21,22,23,24,25] (Additional file 1: Table S1–S3, Fig. S1).
The previous study indicated that the in-hospital mortality among patients with CS infection was 4.73% and with CR infection was 6.77% . For this in-mortality rate, it was calculated that 2041 CR-infected and 2041 CS-infected patients had to be included in the analysis to obtain 80% statistical power to reject the null hypothesis that CS and CR in-hospital mortality rates are equal (Additional file 1: Appendix S1).
All analyses were performed using SAS version 9.4 statistical software (SAS Institute Inc., Cary, NC, USA) and R version 4.1.3 (R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/) .