Mortality risks associated with empirical antibiotic activity in Escherichia coli bacteraemia: an analysis of electronic health records

Abstract Background Reported bacteraemia outcomes following inactive empirical antibiotics (based on in vitro testing) are conflicting, potentially reflecting heterogeneity in causative species, MIC breakpoints defining resistance/susceptibility, and times to rescue therapy. Methods We investigated adult inpatients with Escherichia coli bacteraemia at Oxford University Hospitals, UK, from 4 February 2014 to 30 June 2021 who were receiving empirical amoxicillin/clavulanate with/without other antibiotics. We used Cox regression to analyse 30 day all-cause mortality by in vitro amoxicillin/clavulanate susceptibility (activity) using the EUCAST resistance breakpoint (>8/2 mg/L), categorical MIC, and a higher resistance breakpoint (>32/2 mg/L), adjusting for other antibiotic activity and confounders including comorbidities, vital signs and blood tests. Results A total of 1720 E. coli bacteraemias (1626 patients) were treated with empirical amoxicillin/clavulanate. Thirty-day mortality was 193/1400 (14%) for any active baseline therapy and 52/320 (16%) for inactive baseline therapy (P = 0.17). With EUCAST breakpoints, there was no evidence that mortality differed for inactive versus active amoxicillin/clavulanate [adjusted HR (aHR) = 1.27 (95% CI 0.83–1.93); P = 0.28], nor of an association with active aminoglycoside (P = 0.93) or other active antibiotics (P = 0.18). Considering categorical amoxicillin/clavulanate MIC, MICs > 32/2 mg/L were associated with mortality [aHR = 1.85 versus MIC = 2/2 mg/L (95% CI 0.99–3.73); P = 0.054]. A higher resistance breakpoint (>32/2 mg/L) was independently associated with higher mortality [aHR = 1.82 (95% CI 1.07–3.10); P = 0.027], as were MICs > 32/2 mg/L with active empirical aminoglycosides [aHR = 2.34 (95% CI 1.40–3.89); P = 0.001], but not MICs > 32/2 mg/L with active non-aminoglycoside antibiotic(s) [aHR = 0.87 (95% CI 0.40–1.89); P = 0.72]. Conclusions We found no evidence that EUCAST-defined amoxicillin/clavulanate resistance was associated with increased mortality, but a higher resistance breakpoint (MIC > 32/2 mg/L) was. Additional active baseline non-aminoglycoside antibiotics attenuated amoxicillin/clavulanate resistance-associated mortality, but aminoglycosides did not. Granular phenotyping and comparison with clinical outcomes may improve AMR breakpoints.


Introduction
Antimicrobial resistance (AMR) has received substantial attention for its current and projected threats to safe healthcare worldwide. 1,2 In high-income countries, Gram-negative bacteria, predominantly Escherichia coli, are the leading cause of community-onset bacteraemia, with rising rates of AMR. 3,4 Timely and effective antibiotic therapy for bacteraemia and sepsis is associated with significant mortality reductions. 5, 6 Empirical therapy, therefore, is typically broad-spectrum, although avoiding unnecessarily broad antibiotic exposure is central to mitigating the spread of AMR and other complications, including Clostridioides difficile infection. 7,8 These competing tensions must be balanced while laboratory antibiotic susceptibility results are awaited. Various terminologies are used to describe the situation where antibiotic(s) given to a patient are resistant on susceptibility testing: discordant, inappropriate, or inactive. We use the term 'inactive' throughout, but it should be remembered that this is inactive in vitro, rather than necessarily reflecting no activity in patients. Studies investigating outcomes following inactive therapy have shown contrasting results, potentially due to methodological heterogeneity, including different times to definitive rescue antibiotic treatment, inconsistent definitions of discordant/inappropriate/inactive therapy, 9,10 and different MIC breakpoints defining resistance/susceptibility 11 (amongst others 5, 6 ). Some studies include many bacterial taxa, 9,12 whilst others focus on specific species, often limiting sample sizes. 9,10,13,14 Two meta-analyses found inappropriate antibiotic therapy was associated with increased mortality in sepsis, 5,6 whereas studies specifically of Gram-negative bacteraemia observed no overall association. 15,16 A recent, multicentre US study showed discordant empirical antibiotic therapy for bacteraemia increased in-hospital mortality overall, and varied in frequency across bacterial species (from 5% in β-haemolytic streptococci to 45% in Enterobacterales), as did mortality following discordant therapy, which was highest for Staphylococcus aureus, with only moderate evidence of association for Enterobacterales. 12 Inactive therapy is usually defined using antibiotic susceptibility breakpoints, which are set by expert committees, including EUCAST 17 and CLSI. 18 Wild-type distribution of MICs, pharmacokinetic/pharmacodynamic data and clinical outcomes are all considered, but outcome data can be limited. 19 The advent of 24 h microbiology laboratories (with semiautomated culture/susceptibility testing platforms) and proactive on-call infection consult services potentially reduces the time to effective therapy in those started initially on inactive empirical therapy. This may mitigate any associated harms and influence the necessary breadth of empirical cover. Therefore, we investigated the effect of (in)active, empirical, antibiotic therapy on 30 day mortality following E. coli bacteraemia in patients admitted to our hospital group in Oxfordshire, UK, where these service improvements have been in place for several years.

Methods
We included adults (≥16 years old) with ≥1 blood culture growing E. coli during an admission at Oxford University Hospitals, a large UK teaching hospital group, with four hospitals and 1000 beds, serving a population of 650 000 and providing specialist referral services.
We included the first positive blood culture per patient per 90 day period and only patients who received amoxicillin/clavulanate (also known as co-amoxiclav, a β-lactam/β-lactamase inhibitor combination of amoxicillin and clavulanate widely used in the UK) within their baseline antibiotic regimen. Amoxicillin/clavulanate was the hospital group's firstline antibiotic for suspected sepsis, complicated urinary tract infection, moderate/severe community-acquired pneumonia and intra-abdominal infection. Hospital guidelines also recommended additional single-dose gentamicin in patients with sepsis features to cover potential amoxicillin/clavulanate-resistant infections while results were awaited. We also performed a sensitivity analysis, which included each patient only once, considering only the latest bacteraemia episode per patient.
Antibiotic susceptibility was performed using BD Phoenix automated broth microdilution (or disc diffusion when unavailable), following EUCAST guidelines and breakpoints. 17 Thirty-day all-cause mortality was determined using hospital records that are updated with national data on all deaths. 20 Follow-up was censored at the earliest of 30 days or the last day the patient was known to be alive from national data or hospital records.
We defined the baseline antibiotic regimen as all IV or oral antibiotics administered in hospital within −12 to +24 h of blood collection for each index E. coli-positive culture. Data on antibiotics given in the community prior to admission were not available. We excluded episodes where recorded inpatient antibiotics were commenced .24 h after the index culture.
We included two factors to account for other baseline antibiotics: (i) additional active aminoglycosides; and (ii) additional active 'other' antibiotics (i.e. neither amoxicillin/clavulanate nor aminoglycoside). We considered aminoglycosides separately as these were typically given as a single additional dose, whereas other antibiotics were generally prescribed for longer. To allow for differing effects of additional antibiotics in patients receiving active or inactive amoxicillin/clavulanate, we used four mutually exclusive categories: (i) active amoxicillin/clavulanate (regardless of other drugs); (ii) inactive amoxicillin/clavulanate alone; (iii) inactive amoxicillin/clavulanate with active aminoglycoside only; and (iv) inactive amoxicillin/clavulanate with 'other' active antibiotic. There were insufficient data to include these partial interactions in the categorical MIC model so only main effects for additional antibiotics were included.
We also adjusted for additional baseline factors, including patient characteristics, vital signs, blood tests and hospital factors (see Supplementary methods, available as Supplementary data at JAC Online). Index blood cultures taken within the first 48 h of admission were considered community-acquired and those obtained subsequently nosocomial.
We used multivariable Cox regression, accounting for non-linearity of continuous factors, to model time from index blood culture to death within 30 days, including antibiotic exposures irrespective of statistical significance and other factors based on backwards elimination (exit P . 0.05; see Supplementary methods).
De-identified data were obtained from the Infections in Oxfordshire Research Database, 21 which has approvals from the South Central -Oxford C Research Ethics Committee (19/SC/0403), the Health Research Authority and the national Confidentiality Advisory Group (19/CAG/0144).

Impact of amoxicillin/clavulanate resistance
Using EUCAST breakpoints, adjusting for age, BMI, immunosuppression, prior hospitalization, baseline clinical specialty, blood tests and vital signs (Table 3, Figure 3), there was no evidence    Figure 3).

Other risk factors for mortality
Multiple other factors were independently associated with higher 30 day mortality (Table 3): older age, lower BMI, immunosuppressive condition(s) and prior hospitalization; baseline vital signs, including being unresponsive or responsive only to verbal/painful stimuli, supplementary oxygen; baseline blood test results, including lower albumin, higher alkaline phosphatase, higher urea, lower monocytes, lower neutrophils and higher immature granulocytes; and the baseline clinical specialty (greater mortality associated with acute and general medicine versus medical subspecialty). Higher oxygen saturation was associated with increased mortality, possibly reflecting supplementary oxygen given to the most unwell patients. Since urinary sources were most common, and amoxicillin/clavulanate is predominantly renally excreted, we assessed the impact of additionally adjusting for this in the final high-level amoxicillin/clavulanate resistance model; urinary source was not associated with mortality (P = 0.87) and the estimates of effect of amoxicillin/clavulanate MIC and additional antibiotic therapy were similar.

Sensitivity analyses
In a sensitivity analysis, there was no evidence of an association between mortality and amoxicillin/clavulanate MIC . 16/2 mg/L [aHR versus amoxicillin/clavulanate with MIC ≤ 16/2 mg/L = 1.52 (95% CI 0.93-2.48; P = 0.092), Table S4]. Similar results were obtained when restricting our dataset to only include each patient once (Table S5).

Discussion
In patients with E. coli bacteraemia, after accounting for demographics, comorbidities, past hospital exposure and illness acuity, high-level resistance to baseline empirical amoxicillin/clavulanate (MIC . 32/2 mg/L) was associated with increased 30 day mortality. However, we found no evidence that amoxicillin/clavulanate resistance defined using EUCAST breakpoints was associated with mortality (including MICs of 16/2-32/2 mg/L). This disparity between outcomes and antibiotic breakpoints potentially explains some of the heterogeneity in previous studies investigating the impact of AMR in E. coli bacteraemia.
Our findings are broadly consistent with a recent US study of 26 036 patients with bacteraemia, 12 where there was borderline evidence of increased mortality in patients with Enterobacterales bacteraemia who received inactive empirical antibiotic therapy [OR = 1.23 (95% CI 1.00-1.52; P = 0.054)]. However, this study did not examine the impact of MIC. Another study of Enterobacterales bacteraemia highlighted the methodological differences between CLSI and EUCAST for determining amoxicillin/clavulanate MICs, and found no association between MIC and mortality, by either method and using a variety of breakpoints, but had only limited power with 202 E. coli cases. 11 Mortality risks with antibiotics in E. coli bacteraemia In addition to highlighting adverse outcomes from AMR, our findings suggest that breakpoints for amoxicillin/clavulanate may be set too conservatively, at least for some bloodstream infections. Interestingly, the EUCAST amoxicillin/clavulanate breakpoint for Enterobacterales causing uncomplicated urinary tract infection is .32/2 mg/L, i.e. the level of resistance we found was associated with increased mortality. As urinary tract infection was the most common, presumed bacteraemia source in our study, it is possible that urinary excretion of amoxicillin/clavulanate, coupled with high blood concentrations from IV administration, were sufficient to treat infections with MICs of 16/2-32/2 mg/L. Our approach highlights more generally the benefit of large-scale electronic health records as a tool for reviewing and setting antibiotic breakpoints, with a greater focus on patient outcomes than has previously been possible.   Figure S1 0.005 similar to Figure S1 ,0.001 Yoon et al.
Additional active baseline empirical non-aminoglycoside antibiotics, predominantly cephalosporins and carbapenems, negated the impact of high-level amoxicillin/clavulanate resistance. These additional antibiotics were started prior to microbiology results becoming available, typically replacing amoxicillin/clavulanate, for example, following senior medical review, or transfer from the emergency department to the admitting speciality. There was evidence that these antibiotics as a group were beneficial, i.e. we did not assess variation between cephalosporins versus carbapenems.
In contrast, higher mortality with high-level amoxicillin/clavulanate resistance was similar whether additional active aminoglycoside (majority single-dose gentamicin) was given or not. As local guidelines recommended single-dose gentamicin in patients with high-risk sepsis features, this may simply be a marker that patients who received gentamicin were more unwell. However, we adjusted for illness severity using vital signs and laboratory tests, meaning this is unlikely to be the full explanation. Another possible explanation is that local guidelines recommended gentamicin doses of 3-5 mg/kg to minimize toxicity, in contrast to higher doses, ≥7 mg/kg, that may be required to achieve adequate peak concentrations, particularly among critically ill patients. 22,23 However, our findings may indicate that single-dose aminoglycoside is insufficient to rescue patients with E. coli bacteraemia highly resistant to amoxicillin/clavulanate. Data on the clinical impact of one-off aminoglycosides in E. coli sepsis is sparse, although support exists for its use in less severe infections, e.g. cystitis. 24 In general, combining aminoglycosides with a β-lactam in sepsis does not reduce mortality, but does increase nephrotoxicity. 25 Similarly, whilst aminoglycosides are widely used as an adjunct in neutropenic sepsis with an antipseudomonal β-lactam, this is not backed by guidelines or studies. 25,26 A study strength is that we adjust for confounding more completely than in many other studies, e.g. we account for prior healthcare exposure and other factors that increase the risk of AMR and may also be associated with adverse outcomes. We also used both vital signs and laboratory tests to ensure differences in illness severity at initial presentation were robustly accounted for. Consequently, we found multiple other independent associations with increased mortality: greater age, lower BMI (potentially reflecting lower physiological reserve), prior hospitalization and the presence of immunosuppression, consistent with existing evidence. 9,12,[27][28][29][30][31] Haematology and biochemistry tests, unavailable to the same extent in other recent studies, 9,11,12,32 were strongly associated with mortality. Hypoalbuminaemia, raised alkaline phosphatase and uraemia were associated with increased mortality. [33][34][35][36][37] Neutropenia can be a risk factor or a consequence of sepsis; in this cohort, as amoxicillin/clavulanate was chosen as the primary antibiotic, patients were not believed a priori to have neutropenic sepsis (as local guidelines recommended piperacillin/tazobactam for this); therefore, most of the observed neutropenia, associated with increased mortality, is likely to be a consequence of acute infection. Low monocytes were associated with increased mortality, as previously described. 38 Associations between specific white cell lineages and infection outcomes are also seen in other infections, e.g. eosinopenia and C. difficile, 39 and basophils and eosinophils in COVID-19. 40 Amongst vital parameters, alertness level, oxygen saturation and the recorded use of supplementary oxygen were also associated with mortality.
Other study strengths include its focus on the impact of AMR in E. coli bacteraemia, the most common Gram-negative pathogen, thereby highlighting organism-specific associations with mortality; this study is also one of the largest to assess the clinical impact of EUCAST breakpoints versus alternative definitions of 'active'  Table 3.
Mortality risks with antibiotics in E. coli bacteraemia empirical antibiotic therapy for this pathology. We mitigated some limitations of previous studies: heterogeneity of empirical antibiotic choice by selecting patients administered at least baseline amoxicillin/clavulanate; ambiguity of 'inappropriate' therapy by using in vitro susceptibility-based definitions to define 'active' versus 'inactive' therapy; and assessing the MIC continuum of susceptibility.
Limitations include the lack of data on community-prescribed antibiotics prior to hospital admission, in particular because AMR potentially contributes to failure of community treatment, need for hospitalization and illness severity at admission. Since some of our model variables (e.g. vital signs, blood tests) capture illness severity at admission, we may have underestimated the overall association between MIC and mortality by adjusting for factors that may mediate the pre-hospital impact of AMR. There were relatively fewer patients with MICs of 32/2 mg/L, which limited the power of our study to detect differences in mortality at this MIC; however, point estimates from the categorical model suggested that it was MICs . 32/2 mg/L specifically that were associated with increased mortality rather than 16/2 and 32/2 mg/L. We did not further quantify MICs . 32/2 mg/L, given the limits of the assay used. This was a study of electronic health records from a single hospital group serving a relatively ethnically homogeneous population, with implications for generalizability. Another limitation is that the study antibiotic is a combination of two drugs (amoxicillin and clavulanate) with variation between EUCAST and CLSI approaches to susceptibility testing, with the former using a fixed clavulanate concentration and the latter a fixed ratio, which has implications for which isolates are reported as resistant. 41 Furthermore, in the UK, amoxicillin/clavulanate is the leading β-lactam/β-lactamase inhibitor with an amino-penicillin in use, with a maximum dose of 3 g of amoxicillin component per 24 h. Underdosing of the β-lactam component is another possibility to account for higher mortality seen. A similar β-lactam/β-lactamase inhibitor is ampicillin/sulbactam, which is not marketed in the UK, but can be given in doses of up to 8 g for the ampicillin component in 24 h.
Further work should aim to assess the clinical impact of more granular AMR phenotypes in other causes of sepsis, including other Enterobacterales species, whilst adjusting for severity of illness, capturing both chronic and acute patient contexts. It may also be that bacterial genotypes, i.e. specific resistance mechanisms, are also important in determining outcomes. 42 Identifying which patients are most at risk of inactive empirical treatment, hence most at risk of adverse outcomes, could potentially improve patient outcomes. In summary, high-level (.32/2 mg/L) amoxicillin/clavulanate resistance is associated with increased mortality from E. coli bacteraemia. Disentangling the heterogeneous impact of AMR on mortality may require an organism-specific approach. High-quality electronic healthcare record studies, coupled with more granular resistance phenotyping and genotyping, may improve antibiotic resistance breakpoint setting and potentially in the future lead to clinical guidelines based on MICs and specific patient factors, which in turn may improve outcomes for patients.