This retrospective cohort study was performed in two large academic centres in central Europe, the Jena University Hospital (JUH) in Germany and the Medical University of Vienna (MUW) in Austria. JUH is a 1400-bed academic hospital in the state Thuringia in Germany [19]. MUW is a 1700-bed academic hospital in the city and state Vienna in Austria.

After ethical approval, we included all adult patients with at least one blood culture (BC) positive for E. coli who received an antibiotic monotherapy for at least 72 h between January 2014 and June 2020. Exclusion criteria were polymicrobial bacteraemia, death within the first 72 h, treatment duration of initial antibiotic therapy < 72 h, ESBL-detection, AMP/SLB-sensitive E. coli not treated with PIP/TAZ, in vitro resistance to PIP/TAZ, no hospital admission, initial oral antibiotic treatment, or therapy without a curative goal.

Two different BC systems were used in this study: the BD BACTEC™ FX system (BD Diagnostics, Heidelberg, Germany) at JUH and the BacT/Alert® 3D system (bioMérieux, Marcy l’Etoile, France) at MUW. In general, one to three BC sets per blood draw were collected. The BC bottles were incubated for up to 5 days (JUH) or up to 7 days (MUW) in the respective BC system. Species identification was routinely performed the next day on culture-grown colonies by Vitek® MS (MALDI-TOF (Matrix Assisted Laser Desorption IonizationTime-of-Flight) Mass Spectrometry, bioMérieux, Nürtingen, Germany) or MALDI Biotyper® (MALDI-TOF MS, Bruker, Germany). Antimicrobial susceptibility testing was performed by Vitek-2 system (bioMérieux) and evaluated according to the EUCAST criteria. Only E. coli strains with an AMP/SLB or PIP/TAZ MIC of ≤ 8 mg/l were considered as susceptible to the respective substance. The ESBL status was primarily determined by Vitek-2 ESBL test at JUH and was evaluated by manual ESBL tests (ß LACTA™ test, BioRad; disc diffusion test with an ESBL inhibitor, Mast) at MUW.

The following data were collected from the patients’ medical records: demographic data, comorbidities, Pitt bacteraemia score, implanted devices, source of E.coli BSI (urinary, biliary and other intra-abdominal, vascular catheter-related, respiratory, skin/soft tissue, bone/joint), place of infection (hospital-acquired vs. other place), stay at an intensive care unit (ICU) or intermediate care (IMC) at the onset of BSI, fever > 38 °C, leukocytosis > 12 Gpt/l, Sequential Organ Failure Assessment (SOFA) score at baseline (day of sampling of the first positive BC), 72 h, and 14 days after sampling of the initial positive BC and start of active treatment, duration of BSI (if follow-up BCs were available), first- and second-line antibiotic therapy, duration of antibiotic treatment, source control, length of hospital stay and discharge mode (deceased, alive).

The primary outcome of the study was early treatment response 72 h after the sampling of the initial positive BC and the start of active antimicrobial treatment. Early treatment response was defined as a composite measure and required all of the following parameters: absence of SOFA score increase in patients admitted to ICU or IMC, resolution of fever (temperature < 38 °C), leukocytosis (white blood cell count < 12 Gpt/l), and microbiologic resolution (no documented persistent bacteraemia ≥ 72 h) [20,21,22].

The secondary outcomes included the clinical success 14 days after sampling of the initial positive BC, the 28-day mortality rate, relapsing BSI within 60 days in patients with follow-up BCs and length of hospital stay.

Sample size was based on a prior randomized controlled study in patients with BSI caused by ceftriaxone-resistant but PIP/TAZ-sensitive E. coli or Klebsiella pneumoniae [20]. In this study, clinical response was achieved at median day 3 in the PIP/TAZ group (IQR 1–5 days), indicating a proportion of 50% cases with a clinical response within 72 h. Considering these results, a sample size of 103 per group is needed to detect an absolute difference of 20% (early treatment response: 50% vs. 70%) on a significance level of 0.05 and 80% power with a two-sided Fisher’s exact test. Therefore, in this study the minimum total sample size is 309 (3 × 103) adult patients with E. coli BSI.

To reduce the expected bias between the three patient groups, propensity score matching was performed. Propensity scores were calculated using a logistic regression model with patient group as the dependent variable, and variables potentially associated with the treatment decision and the clinical response (sex, age and centre) as the independent variables. In the first step, we calculated propensity scores for a subset containing cases from patient groups A and C, with group A considered the treatment group and group C as the control group. We matched group C to group A at a 1:1 ratio using the K-nearest-neighbors method without replacement. In the second step, group B was matched to the pre-matched group A, by repeating the matching procedure. Propensity score matching was performed using the R-package MatchIt (Ho, 2011) (supp. Fig. 1).

Baseline demographics were assessed using descriptive statistics. Continuous and discrete numerical variables were reported as median and interquartile range (IQR) or median and range. Categorical variables were reported as frequency and percentage. Fisher’s exact test was used to compare the primary and secondary endpoints between the three groups. p values were adjusted for multiple testing using the Holm–Bonferroni method. Univariate logistic regression was used to model the early clinical response based on the therapy group. A multiple logistic regression model was implemented to assess the influence of the antibiotic therapy on early clinical response in patients with BSI caused by AMP/SLB-resistant E. coli. This model included antibiotic therapy, ICU or IMC admission, centre, Pitt bacteraemia score, neutropenia, liver disease and BSI source. A two-sided significance level of 0.05 was applied in all models.

Statistical analysis was performed using R version 4.1.3 (R Core Team (2022). Vienna, Austria). The R packages ggplot2 (Wickham, 2016), ggpubr (Kassambara, 2020) and viridis (Garnier, 2021) were used for graphical representation of the data.