Cefiderocol for treating severe aerobic Gram-negative bacterial infections: technology evaluation to inform a novel subscription-style payment model

Susceptibility studies, pharmacokinetic/pharmacodynamic studies and breakpoints

Susceptibility studies are in vitro studies that report the results of AM-susceptibility testing (AST). AST is a laboratory method where isolates taken from patients (from infections or during screening) are grown in vitro (cultured) and tested for their susceptibility to various AM treatments. The AM being tested is applied at increasing concentrations to separate cultures of the sampled isolate, and the degree to which microbial growth is inhibited at each concentration is assessed. The lowest concentration at which microbial growth is inhibited is known as the minimum inhibitory concentration (MIC).

Clinical breakpoints distinguish between isolates where there is a likelihood of treatment success from those where treatment is more likely to fail. 24 If the MIC of a given isolate is at or below the breakpoint, the isolate is judged to be ‘susceptible’ (S). If it is above the breakpoint, the isolate is judged to be ‘resistant’ (R). For some AMs, there is also an intermediate category (I), which more recently has become ‘susceptible – increased exposure’ indicating that a higher dose of the drug should be used to elicit a response. They may also report the concentration at which 50% of isolates were inhibited (MIC 50), and the concentration at which 90% were inhibited (MIC 90).

The methods for setting breakpoints are not standardised. Currently, they are generally set by considering:24

• the pharmacokinetic (PK) data: how the body affects the drug with respect to absorption, distribution, metabolism and excretion, usually obtained from studies in healthy volunteers

• the pharmacodynamic (PD) data: how the drug affects the body (efficacy and toxicity) at its site(s) of action, usually obtained from in vitro studies, hollow fibre studies, animal studies and human studies. These data are used to set PD targets, for example for time above MIC

• mathematical models (e.g. Monte Carlo simulation) to assess the likelihood of achieving the targets suggested by the PD data

• any available clinical data linking treatment to clinical outcomes (e.g. from RCTs or observational studies).

Pharmacokinetic/pharmacodynamic studies are conducted to estimate how much drug will be available at the site of interest, and for what period of time at a given dose. One of its primary uses is by manufacturers and regulatory bodies to decide on the appropriate dose and dose frequency of the drug, such that it is likely to be available at concentrations that are likely to have an effect at the sites of interest.

There are two main organisations that set breakpoints, the Clinical Laboratory Standards Institute (CLSI) in the USA and the European Committee on Antimicrobial Susceptibility Testing (EUCAST) in Europe. These two organisations use different methodologies to set breakpoints, leading to differences in the breakpoints set both in absolute and relative terms, between treatments. They also describe different laboratory methods to assess MICs. In addition, many labs may use commercial assays, conducted according to manufacturer’s instructions. Clinical advisors to EEPRU indicated that it was unclear to what extent CLSI, EUCAST and commercial methods would produce the same absolute values, and in the event that values were different, whether relative values between treatments would also be different (i.e. the difference in absolute values was not consistent across treatments). In the UK, the British Society for Antimicrobial Chemotherapy (BSAC) now recommends use of EUCAST methods and breakpoints.

Susceptibility studies tend to report the proportion S, I and R, or list the number of isolates at each MIC. An example is given in Table 3. Here, for cefepime, the breakpoint is 1 mg/l, and since all isolates had MICs higher than the breakpoint, none were susceptible. For cefiderocol, with a breakpoint of 8 mg/l, 90.9% were susceptible, since only one isolate had a MIC above this point.

Studies reporting the susceptibility of MBL Enterobacterales or P. aeruginosa isolates to cefiderocol and at least one comparator

Since no RCTs or observational evidence were identified that met the requirements of the HVCSs, EEPRU considered the evidence relating to in vitro susceptibility. EEPRU’s approach is supported by the company submission,82 which states that ‘the in vitro data for cefiderocol should be considered the primary source of evidence to support the effectiveness and positioning of cefiderocol and is actually the only source of evidence for the risk-based empiric setting’.

Four susceptibility studies met the inclusion criteria of the review. 34,35,78–81 The characteristics of these studies are presented in Table 7. The three RCTs13–15 were also examined for relevant susceptibility data, but none were relevant to the susceptibility synthesis as they did not present data by mechanism. It is noted that the Shionogi company submission82 recommends PK/PD evidence should be used alongside clinical and susceptibility data to extrapolate to other infection sites. However, the clinical advisors to EEPRU advised that as cefiderocol and comparators penetrate to the sites of interest, it was reasonable to link directly between susceptibility and clinical outcomes (see Table 6). Furthermore, neither the Shionogi company submission nor discussion with the clinical advisors to EEPRU identified a quantitative approach to linking PK/PD evidence to clinical outcomes. Consequently, as part of the mapping exercise, PK/PD studies were excluded.

Policy Research Unit in Economic Methods of Evaluation in Health and Social Care Interventions first considered whether any one of the studies met all the requirements of the evaluation (ideally consecutive English data from a multisite study reporting outcomes for all relevant comparators, using BSAC/EUCAST breakpoints and laboratory methods), and could fulfil the evidence needs of the project without the need of a meta-analysis. Since the data received from PHE in response to EEPRU’s request did not report any evidence for cefiderocol, these data were not able to fulfil the requirements of the project, and a meta-analysis of susceptibility studies was planned. There were also a number of other limitations to the PHE data. Isolates have not historically been routinely submitted by testing centres which may limit how representative these data are of the true distribution of MBL susceptibilities in England. In addition, there is inconsistency in the testing methodologies used by local laboratories (albeit the majority use EUCAST). 93 Finally, not all isolates were tested for each comparator, and a compromise had to be made in conducting the analysis whereby to preserve internal validity only isolates tested among all comparators were included (see Report Supplementary Material 1) which may have introduced selection bias. Given this approach, we excluded PHE evidence for fosfomycin as this would have resulted in only 18 isolates being available for analysis across both MBL Enterobacterales and P. aeruginosa populations.

There was some uncertainty during the development of the PICOS about whether meropenem should be considered a relevant comparator. This led to inconsistencies in the inclusion of evidence relating to meropenem. Data from one study (Kohira et al. 201680) and the meropenem data from PHE were excluded erroneously from the synthesis. The sensitivity analyses performed to address this are described in Statistical synthesis plan.

Limitations of the data available from the published study reports

Two of the studies, Longshaw et al. (2020)81 and Kazmierczak et al. (2019),79 drew isolates from SIDERO CR and SIDERO WT (wild type), respectively. These were surveillance studies conducted by Shionogi. The two further studies78,94 identified by the review were also funded by Shionogi.

Policy Research Unit in Economic Methods of Evaluation in Health and Social Care Interventions identified limitations with all four studies with respect to their relevance to the synthesis being conducted to inform estimates of relative effectiveness in the HVCS:

• Kazmierczak et al. (SIDERO WT)79 only reported MIC50, MIC90 and range, not % susceptibility. While these metrics could have been used to reconstruct the distribution curves and apply a breakpoint to generate an estimated % susceptibility, this was thought to introduce too much uncertainty to the estimates. In addition, not all three metrics were reported for all relevant subgroups in Kazmierczak et al. (SIDERO WT). 79

• Data from SIDERO CR [Longshaw et al. (2020)]81 covered only Europe, whereas the full data set used in Shionogi’s response to a data request from EEPRU dated 14 June 2021 (see Report Supplementary Material 1) used worldwide data. After consultation with our clinical advisers, it was decided that, ideally, we would include worldwide data in the synthesis since resistance mechanisms frequently arrive in the UK from elsewhere (see Table 6).

• Neither Johnston et al. (2020)78 nor Kohira et al. (2016)80 reported susceptibility using EUCAST breakpoints, only CLSI breakpoints.

• All four sources of data used CLSI laboratory methods, which may not be equivalent to EUCAST laboratory methods.

• None of the studies included data for fosfomycin.

Shionogi was approached to provide additional data to fulfil these requirements (data request dated 11 August 2021, see Report Supplementary Material 1) for all four studies. The request was for worldwide data, reporting percentage susceptibility using EUCAST and separately using CLSI breakpoints for cefiderocol and all relevant available comparators (in case data for fosfomycin was unpublished, but held by Shionogi). Shionogi supplied the requested analyses for SIDERO CR35 and SIDERO WT34 (see Table 7 below). The data were reported for Enterobacterales and P. aeruginosa separately, and included isolates with carriage or co-carriage of MBLs. The data were not restricted by carbapenem sensitivity, or any other sensitivity or phenotype, and counted intermediate susceptibility as resistant, in accordance with EEPRU’s requirements. Data were reported using CLSI and EUCAST breakpoints, but using CLSI laboratory methods. No data for fosfomycin were available. The company was unable to provide additional data for Johnston et al. (2020)78 or Dobias et al. (2017)67 as they did not have access to the raw data. Therefore, Dobias et al. (2017)67 was excluded from the syntheses, while Johnston et al. (2020)78 could only be included in a synthesis of data using CLSI breakpoints. Consequently, three studies (SIDERO CR, SIDERO WT and Johnston 2020)78 of Cefiderocol were eligible for inclusion in the statistical synthesis. Due to the uncertainty around comparators noted above, data for Kohira et al. (2016)80 were not requested, but the data reported in the journal publication were included in the review in the sensitivity analyses conducted to address this issue (see Statistical synthesis plan).

Characteristics of studies entering the network meta-analysis

Across the four cefiderocol studies, the sample size ranged from n = 6980 isolates to n = 34378 isolates. The cefiderocol studies collected isolates internationally, from multisite locations. Expert advice indicated that resistant infections tend to arrive in the UK from around the world, and consequently isolates collected from any location were of relevance to the assessment. All studies were therefore retained in the analysis.

The isolates were collected since 2014 for the SIDERO studies, Johnston et al. (2020)78 included isolates collected over the period 2002–17 and Kohira et al. (2016)80 included isolates from 2000 to 2011. Expert advice indicated that isolates collected since 2012 were of highest (but not exclusive) relevance, so all studies were retained in the analysis.

The PHE data did not meet the inclusion criteria for either the cefiderocol or fosfomycin review since it did not report data for cefiderocol or fosfomycin. However, it was considered the best source of UK data for comparators, which were not well covered by the Cefiderocol review, so the PHE data were included for this reason.

Quality assessment of studies entering the meta-analysis

A summary of the quality assessment is presented in Table 8. No study scored low risk for all items.

Among the cefiderocol studies, CLSI laboratory methods were used but EUCAST breakpoints applied in three of four studies, and CLSI methods and breakpoints were used in the fourth. These issues led to a high risk of bias in ‘Outcome measurement’ compared to practice in England. All four scored unclear risk for ‘Target population’ and ‘Sampling strategy’, largely due to a lack of clarity about how isolates were selected for inclusion in the data request response, or how they were selected for β-lactamase testing from the wider population. All scored low risk for missing data, since all isolates were tested for all comparators.

Among the fosfomycin studies, only one scored low risk for ‘Target population’, while the remainder scored unclear high risk, largely due to a lack of clarity around how isolates were selected for β-lactamase testing, and whether isolates collected before 2012 would impact on susceptibility estimates. All studies scored unclear risk for ‘sampling strategy’, usually because it was not stated how isolates were selected, or where it was reported, it was not clear whether the strategy was appropriate. Outcome measurement was either low or unclear risk, due to a lack of clarity around laboratory methods. All scored low risk for missing data, since all isolates were tested for all comparators.

The PHE data were unclear on all domains, since PHE were unable to ascertain how isolates were tested, it was unclear what criteria led to submission of isolates, and it was unclear whether isolates were selected for testing of some interventions on the basis of suspected resistance or susceptibility, which may introduce selection bias.

Base-case network meta-analysis: metallo-beta-lactamases Enterobacterales infections with European Committee on Antimicrobial Susceptibility Testing breakpoint

Eight studies contributed to the NMA of MBL Enterobacterales infections with EUCAST breakpoints, considering a total of eight comparators, and the full network diagram is shown in Figure 2. Three studies83,84,86 contained 100% susceptibility counts for one or more of the included comparators and therefore had a continuity correction applied prior to synthesis.

The relative susceptibility for each comparator relative to colistin is shown in Figure 3a. The model fitted the data well, with a total residual deviance of 33.50, which was close to the number of data points included in the analysis of 35. The between-study SD was 1.45 (95% CrI 0.93 to 2.35), which indicates extremely high heterogeneity. Cefiderocol was associated with a lower susceptibility relative to colistin (OR 0.32, 95% CrI 0.04 to 2.47), but the result was not statistically significant. Cefiderocol was also associated with an 11% probability of being the most effective treatment; median rank 3. The remainder of the treatments were also associated with lower susceptibility than colistin, but the results were not statistically significant. For all comparators, the high between-study SD results in wide 95% PrI. The updated analysis including the missing PHE meropenem data were very similar for cefiderocol (OR 0.31, CrI 0.04 to 2.18) and comparators.

Inconsistency checking was performed using the UME model. The model fitted the data well and the DIC was similar to the base-case NMA model (Appendix 4, Summary of network meta-analyses, Table 43). The estimated between-study SD is slightly smaller from the UME model compared to the base-case NMA model, but it still indicates extremely high heterogeneity (1.08, 95% CrI 0.67 to 1.82). The deviance plot (see Appendix 4, Inconsistency checks, Figures 38 and 39 show the deviance plots used to assess inconsistency) shows no obvious improvement in fit when using the UME model. As there is no evidence of inconsistency, no further steps were taken.

FIGURE 3.

Forest plot of OR vs. colistin for MBL Enterobacterales with EUCAST breakpoint (SIDERO, fosfomycin and PHE studies). (a) all comparators, missing PHE data; (b) All comparators, including missing PHE data; and (c) only comparators specific to the pathogen. PB, probability of being the best treatment.

Base-case network meta-analysis: metallo-beta-lactamases P. aeruginosa infections with European Committee on Antimicrobial Susceptibility Testing breakpoint

Four studies contributed to the NMA of MBL P. aeruginosa infections with EUCAST breakpoints for SIDERO, fosfomycin and PHE studies, considering a total of four comparators, and the full network diagram is shown in Figure 4. All four studies (SIDERO WT data request,34 SIDERO CR data request,35 PHE data request and Cuba et al. )88 contained either zero or 100% susceptibility counts for one or more of the included comparators and therefore a numerical adjustment was applied prior to synthesis.

FIGURE 4.

Network diagram of all studies contributing to the NMA (MBL P. aeruginosa with EUCAST breakpoint for SIDERO, fosfomycin and PHE studies).

The relative susceptibility for each comparator relative to colistin is shown in Figure 5. The model fitted the data well, with a total residual deviance of 9.3, which was considered close to the number of data points included in the analysis of 11. The between-study SD was 0.87 (95% CrI 0.04 to 2.76), which indicates high heterogeneity. Cefiderocol was associated with a lower susceptibility relative to colistin (OR 0.44, 95% CrI 0.03, 3.94), but the result was not statistically significant. Cefiderocol was also associated with a 20% probability of being the most effective treatment; median rank 2. The remainder of the treatments were associated with no susceptibility. For all comparators, the high between-study SD results in wide 95% PrI. No inconsistency checking could be performed because the network does not include a feedback loop.

FIGURE 5.

Forest plot of OR vs. colistin for MBL P. aeruginosa with EUCAST breakpoint (SIDERO, fosfomycin and PHE studies). PB, probability of being the best treatment.

Sensitivity analyses

A summary of the results of the sensitivity analyses is presented in this section. Details of the results are presented in Appendix 4, section Sensitivity analysis network meta-analysis results.

Methods

A comprehensive review of the safety of comparators was not possible within the time frame of this evaluation. Adverse events included in the model for the intervention and comparators are described in Model structure for the risk-based empiric setting. Clinical advisors to EEPRU indicated that cefiderocol is predominantly a safe treatment, but that colistin and aminoglycosides have significant adverse events relating to acute kidney injury (AKI). Another key adverse event related to antibiotic use is the emergence of Clostridium difficile in a patient’s digestive tract, which can lead to diarrhoea and serious damage to the colon. EEPRU conducted a review of the RCT trial evidence for cefiderocol to establish whether it supported the clinical view that cefiderocol is a safe treatment. EEPRU were especially interested in establishing safety compared to toxic alternatives (colistin and aminoglycosides) and the other ‘safer’ treatments used in the HVCSs.

Rates of serious treatment-related adverse events, nephrotoxicity adverse events and C. difficile infections were extracted from the included RCT publications and/or their ClinicalTrials.gov national clinical trial record. Only RCTs relating to the sites of interest were reviewed, due to time and resource constraints. Only RCTs were considered as these give comparative data. In the absence of nephrotoxicity events, other kidney and renal adverse events were extracted. The data were then synthesised narratively for any important safety signals.

A pooled analysis of CREDIBLE-CR, APEKS-NP and APEKS-cUTI was presented by Shionogi in their submission to NICE,82 but since comparators in the studies differed, EEPRU preferred to consider the data from each RCT separately.

Results

The extracted adverse event data in the RCTs for cefiderocol are presented in Table 12. Across three RCTs, the proportion of patients with serious adverse events was lower with Cefiderocol compared to: imipenem/cilastatin (4.7% vs. 8.1%; APEKs-cUTI),13 meropenem (2.0% vs. 3.3%; APEKs-NP)14 and BAT [which comprised colistin-based regimens n = 30 (61%); non-colistin-based regimens (n = 19 (39%)] (1% vs. 10%, Bassetti et al. 2019/20 CREDIBLE). 15 A statistical significance for the between-group differences was not reported by any RCT.

None of the included studies reported specifically on nephrotoxicity. Related adverse events included AKI and renal disorder. All three RCTs reported both of these adverse events. The proportions of patients with either AKI or renal disorder was 0.0% for both cefiderocol and the comparator in the two RCTs that did not use colistin in the comparator arm. 13,14 The proportions of patients with AKI and renal disorder were greater with the comparator in the RCT by Bassetti et al. ,15 where colistin was used within the comparator arm (BAT 8.2% vs. cefiderocol 0.0% and 2.0% vs. 0.0%, respectively). A statistical significance for the between-group difference was not reported by any RCT.

One study reported C. difficile infections. The proportion of patients with C. difficile was greater with the comparator (meropenem) compared to cefiderocol (1.3% vs. 0.7%). 14 A statistical significance for the between-group differences was not reported.

Other adverse events were generally very low across the studies, with slightly greater proportions of patients experiencing adverse events with cefiderocol compared to the comparator in all studies. Statistical significance for the between-group differences was not reported.

Population

The patient populations modelled align with the decision problem outlined in Decision problem. These are summarised in Table 14.

Intervention

Cefiderocol is considered as monotherapy only due to a lack of in vivo or in vitro evidence about how it performs in combination with other agents. The clinical advisors confirmed that monotherapy was more likely to be used in practice but indicated that combination therapy was more likely for P. aeruginosa.

Comparators

A wide range of drugs is considered relevant in the HVCSs, and different drugs were considered relevant depending on the site, pathogen, mechanism and setting. The full list of comparators is provided in Population, intervention, comparison, outcomes, study designs for high-value clinical scenarios. Due to the paucity of data available to inform the comparative effectiveness assessment (see Clinical evidence), and our reliance on in vitro susceptibility data to inform comparative effectiveness, it was possible to take a simplified approach to modelling these comparators rather than conducting a fully incremental analysis of all available AM options as is typically recommended in economic evaluation. The approach taken is documented in the following section.

Model structure for microbiology-directed setting

In the MDS, each patient’s susceptibility to available treatment options is known, and treatment can be tailored accordingly. Based on feedback from our clinical advisors, the two main reasons for initiating treatment with cefiderocol (provided patients are susceptible to it) within the MDS HVCSs would be that patients are either: (1) not susceptible to any other available treatment options (i.e. patients are completely MDR to relevant existing treatment options); or (2) that the only other treatments to which they are susceptible carry an elevated risk of nephrotoxicity. We include colistin and aminoglycosides within the category of nephrotoxic drugs as our clinical advisors indicated that they are likely to be associated with elevated levels of nephrotoxicity. To reflect these considerations, patients within the MDS are divided into three categories based on their susceptibility to existing therapies and, within each category, further subdivided according to their susceptibility to cefiderocol. Table 15 shows these subgroups, how they determine treatment choice under existing care and how that would change if cefiderocol was to become available to this patient group. The groups for which a switch to cefiderocol is expected are highlighted in bold.

In the group of patients who are susceptible only to colistin or an aminoglycoside, and susceptible to cefiderocol, cefiderocol offers a safety advantage. In the group of patients who are not susceptible to any available treatment options and, in the absence of the new treatment under evaluation, would receive multidrug salvage therapy, cefiderocol offers a safety and efficacy advantage. This is because, for many patients, multidrug salvage therapy would be expected to include a colistin or aminoglycoside component. Throughout the modelling, isolates classed as intermediate resistant are grouped with those which are resistant as patients infected with intermediate resistant and resistant pathogens are expected to experience similar outcomes in the HVCS based on feedback from EEPRU’s clinical advisors, and much of the data relating mortality and hospitalisation to susceptibility follow this grouping. Our clinical advisors noted that it may be possible to overcome intermediate resistance via higher dosing, but also considered that it would be difficult to evidence this within the model. Given the diverse range of data sources informing susceptibility and the link between susceptibility and outcomes, and the level of reporting within these studies, it was not feasible to explore the implications of differential outcomes between intermediate resistant and resistant patients.

In the MDS, the model is, therefore, driven by the proportion of individuals within each category of ‘susceptibility to existing drugs’ and the proportion of individuals susceptible to cefiderocol. This is based on susceptibility data as documented in section Clinical parameters – susceptibility evidence. The comparison made within the model is between the overall MDS cohort who receive tailored therapy with the new drug available (column four of Table 15) and the overall MDS cohort who receive tailored therapy under existing treatment options only (column 3 of Table 15).

Importantly, the fact that the susceptibility profile is known prior to initiation of treatment in the MDS, alongside the assumption that susceptibility is the sole predictor of treatment effectiveness, means that we do not need to model each individual treatment option within the MDS. For example, it is not relevant (to clinical outcomes) whether a patient is susceptible to fosfomycin or aztreonam as susceptibility to these treatments would be assumed to result in the same outcomes. Although there are differences in the costs of therapies, these are modest in relation to other costs such as that of hospitalisations which may include periods in the ICU/HDU. In practice, patients may receive a combination of agents, but this is not modelled explicitly due to a lack of evidence. EEPRU’s clinical advisors considered it reasonable to assume that, in the MDS, patients susceptible to a single AM within a multiagent regimen perform as well as those susceptible to all components of that regimen (i.e. it does not matter if you are susceptible to drug A, drug B or drug A and B, as long as you are susceptible to one of the agents received).

Following receipt of treatment in the MDS, patients are modelled to experience one of the four alternative 30-day outcomes which determine their long-term outcomes (see Figure 7). A decision tree is used to determine the distribution of patients across these categories at 30 days. This is as shown in Figure 8. Probabilities highlighted in bold differ by treatment in this figure. In the MDS, we only model one line of treatment explicitly, though hospitalisation and mortality evidence will reflect the fact that some patients entering the MDS receive multiple lines of therapy.

FIGURE 7.

Thirty-day outcomes in the MDS.

FIGURE 8.

Decision tree used to calculate impact of AKIs on 30-day outcomes in MDS.

All patients face a risk of death due to their infection and comorbidities (p_bgrdD30d_MDS). The risk differs according to whether patients have received a treatment to which they are susceptible or not. In the MDS, given that treatments are tailored according to patients’ known susceptibility profiles, only patients infected with a fully MDR infection (who receive multidrug salvage treatment) are expected to face the elevated risk of death of those non-susceptible to treatment. The efficacy advantage of cefiderocol is, therefore, driven by the proportion of people who, rather than having ‘multidrug salvage’, are given cefiderocol (see Table 15), as these patients switch from experiencing the mortality of non-susceptible patients to experiencing the mortality of susceptible patients.

In addition, patients face differing drug-related risks of experiencing an AKI. Patients who experience an AKI face an elevated risk of death compared to those who do not. When modelling the effect of AKI on mortality, we account for the fact that the available mortality data already reflect both the underlying risk of AKI associated with currently available non-colistin/aminoglycoside AMs, and the background risk of AKI associated with patients underlying comorbidities and infection (see Figure 8). Patients who experience an AKI and survive until 30 days face a risk of adverse long-term outcomes according to whether they have: (1) recovered their renal function; or (2) suffered irreversible renal failure, that is, developed chronic kidney disease (CKD).

At 30 days, patients who are discharged alive without renal dysfunction are assigned a comorbidity-adjusted QALY outcome estimated using an alive–dead area-under-the-curve approach. This is independent of the assigned treatment, as patients alive at 30 days without a history of AKI are assumed to experience similar outcomes regardless of the treatment they received for their infection.

Patients discharged with recovered renal function face the same HRQoL outcomes, but they face an additional risk of progressing to CKD and elevated mortality. Patients discharged with CKD or who develop CKD face further elevated mortality, reduced HRQoL and additional healthcare costs. The experience of the two groups of patients with a history of AKI is modelled as a semi-Markov process (with transition probabilities dependent on time in model) for all transitions as shown in Figure 9.

FIGURE 9.

Markov model used to calculate post-30-day outcomes in patients with recovered renal function and irreversible renal failure.

Model structure for the risk-based empiric setting

The approach taken in the ES is similar to that taken in the MDS in terms of the possible 30-day outcomes patients can experience and the long-term implications of these outcomes. However, the decision tree describing differences across comparators in the first 30 days is more complex for two reasons. Firstly, there is a need to model outcomes both in those correctly identified as having the pathogen-mechanism combination suspected, as well as those who were labelled as high risk but in fact have a different causative pathogen or mechanism. Secondly, there is a need to model both the ES phase of treatment, and progression of some patients to the MDS for further treatment.

Unlike in the MDS, in the ES the susceptibility of patients to treatments provided is unknown at the time of initiating empiric treatment. It is, therefore, necessary to model the probability of susceptibility to individual treatment combinations as this determines clinical outcomes and, in particular, the need for further treatment. Since, as documented in Population, intervention, comparison, outcomes, study designs for high-value clinical scenarios, there are a number of feasible treatment combinations for these patients, to simplify the modelling we compare empiric use of cefiderocol in the ES with two alternative treatment options:

1. The non-colistin or aminoglycoside-based treatment combination with the current highest estimated susceptibility in the UK population.

2. The colistin or aminoglycoside-based treatment combination with the current highest estimated susceptibility in the UK population.

When considering possible treatment pathways in the ES, the following possible pathways are relevant (shorthand labels: E1, E2nca, etc. are used in the results section):

1. E1: empiric use of cefiderocol followed by existing treatments in the MDS.

2. E2nca: empiric treatment using non-colistin or aminoglycoside-based existing therapies followed by existing therapies in the MDS.

3. E2ca: empiric treatment using colistin or aminoglycoside-based existing therapies followed by existing therapies in the MDS.

4. E3nca: empiric treatment using non-colistin or aminoglycoside-based existing therapies followed by cefiderocol use if indicated in the MDS.

5. E3ca: empiric treatment using colistin or aminoglycoside-based existing therapies followed by cefiderocol use if indicated in the MDS.

Repeated usage of cefiderocol in the MDS for patients who fail cefiderocol in the ES was not modelled as this was not considered to represent a priority use for cefiderocol.

Thirty-day outcomes in the ES are determined by a decision tree which comprises three subcomponents:

1. the risk of carrying the pathogen mechanism of concern

2. outcomes at the point at which patients are assessed for MDS treatment, that is at around 3–5 days when susceptibility results report and

3. 30-day outcomes following MDS assessment.

Each of these is considered in more detail below.

1. Risks of carrying the pathogen mechanism of concern.

Patients may or may not have the suspected pathogen mechanism of concern. We assume that patients who do not have the pathogen mechanism of interest experience the same effectiveness outcomes regardless of the choice of empiric treatment (though safety differs), as our clinical advisors confirmed that these patients represented a broadly susceptible population (rather than a population enriched with pathogens carrying other resistance mechanisms), and that for this reason effectiveness is likely to be similar across all empiric treatment options considered. For simplicity, we assume that patients who have a different pathogen-mechanism experience the susceptibility associated with colistin/aminoglycoside-based therapy in people with the pathogen mechanism of interest regardless of the choice of treatment. Colistin/aminoglycoside-based therapy was chosen as more representative of outcomes across susceptible patients as this treatment class showed robust and high susceptibility across subgroups and scenarios. The structure of this element of the ES model is presented as Figure 10.

FIGURE 10.

First component of 30-day outcomes model for ES: risk of carrying pathogen mechanism of concern.

1. Outcomes at the point at which patients are assessed for MDS treatment.

At initiation of empiric treatment, patients are classified by the model as susceptible or non-susceptible to their empiric therapy. As in the MDS, susceptibility is the driver of differences in effectiveness across treatments. Note that we are able to model differences in susceptibility across treatments used in the ES dependent on whether a patient is susceptible or non-susceptible, even though clinicians will not observe this information until patients enter the MDS.

At the point at which patients’ microbiology results become available, patients may have died, may require initiation of a new AM treatment (e.g. due to lack of efficacy) or may complete their course of empiric treatment. The probability of these three outcomes depends on whether patients were susceptible to their empiric treatment or not, but not directly on the choice of specific treatment. Patients who have received empirically a treatment to which they are later found not to be susceptible are all assumed to require further treatment in the MDS, provided they survive until microbiology results are available. This assumption is based on evidence presented in Tumbarello et al. (2013), which found that all patients who received inappropriate empiric treatment and survived until their microbiology results were received switched to appropriate therapy (further details on Tumbarello et al. are provided in 8.2.3.4). 125 The structure of this element of the ES is shown in Figure 11.

FIGURE 11.

Second component of 30-day outcomes model for ES: outcomes at the point at which patients are assessed for MDS treatment. NB: Mortality (p_bgrd_Dst_S and p_bgrdDst_nonS) is also adjusted to reflected differences in mortality due to AKI, in the same way as shown in Figure 12, but this is not shown for parsimony.

1. 30-day outcomes following assessment for MDS treatment.

People who survive until the time point of assessment for MDS treatment enter the third part of the decision tree which is shown in Figure 12. Those requiring no further treatment face a risk of dying between this point and 30 days which depends on whether they experienced an AKI. Those surviving to 30 days face the possibility of entering the (1) alive; (2) alive with recovered renal function or (3) alive with irreversible renal failure health states described in Figure 12. While patients may experience an AKI following empiric treatment, clinicians confirmed that in this patient group, where treatment options are limited, the AKI alone would not typically trigger a treatment switch, provided the treatment was effective.

FIGURE 12.

Third component of 30-day outcomes model for ES: 30-day outcomes following assessment for MDS treatment.

Patients who require further treatment enter the MDS component of the model. Their outcomes depend on whether they experienced an AKI following first-line treatment (i.e. this is ‘remembered’ within the model) as this determines both their outcomes (patients who experience an AKI experience elevated mortality and the implications of reversible or irreversible kidney damage) and their choice of treatment in the MDS. Our clinical advisors informed us that patients requiring further treatment in the MDS, who experienced an AKI following treatment in the ES are unlikely to receive colistin- or aminoglycoside-based treatment in the MDS. Patients who fit this profile, and are only susceptible to colistin or aminoglycoside-based treatment are, therefore, assumed to receive multidrug salvage therapy in the MDS, or the new drug if available. For these patients, multidrug salvage therapy is assumed not to include colistin or an aminoglycoside. Instead, they are assumed to receive the outcomes of multidrug salvage therapy without elevated risk of AKI.

In the absence of evidence to support more detailed modelling, we assume that a patient’s susceptibility to treatment is the same in the ES and MDS. In reality, patients entering the MDS who were already assessed as high risk of carrying a highly resistant pathogen in the ES are likely to receive aggressive treatments in the ES which may change their resistance profile in the MDS. The nature of the effects on individual resistance are hard to predict as they are influenced by the treatment received in the ES, the effectiveness of this treatment and the development of acquired resistance. These are not, therefore, considered within the model.

Identification of evidence

Susceptibility evidence and evidence linking susceptibility to mortality and hospitalisation were obtained via the systematic reviews and structured expert elicitation as described in Review results, Statistical synthesis and Review questions 4 and 5. Other key clinical parameters were obtained from existing systematic reviews where possible; otherwise, clinical parameters were obtained from existing UK cost-effectiveness models. Quality-of-life weights (utilities) were obtained from a systematic review (described below) and cost parameters via targeted searches.

Clinical parameters – susceptibility evidence

The susceptibility data used in the model base-case analysis are summarised in Table 16. These represent the mean values of the samples used in the probabilistic sensitivity analysis (PSA), along with 95% percentile-based confidence intervals. Five key susceptibility parameters inform the model:

• One parameter describes susceptibility to cefiderocol in the MDS and ES.

• Two parameters describe susceptibility to colistin/aminoglycoside-based therapy and to non-colistin/aminoglycoside-based therapy in the ES.

• Two parameters describe the number of individuals in each category of susceptibility in the MDS as shown in Table 15 (namely, susceptible to a non-colistin/aminoglycoside AM, susceptible only to a colistin/aminoglycoside AM).

Susceptibility to existing drugs is obtained from both the analysis of PHE data and the NMA, as described in Statistical synthesis. These analyses can be combined to provide evidence on absolute rates of susceptibility to AMs for the HVCS. This evidence required further adjustment before it could be used in the economic model, with different adjustments for the ES and MDS. The methods employed to obtain estimates for these two settings are discussed in turn, with further details provided in Appendix 7. Of note, while evidence on susceptibility to meropenem was available, this was not used in the economic modelling. This is because clinical advice was that, for meropenem, susceptibility among carbapenem-producing pathogens was not a good surrogate predictor of clinical outcomes. This reflects advice in the literature. 126,127 Hence, while meropenem is included as a comparator in the PICOS, it is assumed to have zero efficacy in the economic modelling (and so not actively modelled).

Susceptibility is estimated to be specific to the pathogen-mechanism subgroup of interest but is assumed to generalise across sites and settings. This was considered a reasonable assumption by our clinical advisors and preferable to further subgrouping the susceptibility data given the small sample sizes available to inform these parameter estimates for the HVCS.

Clinical parameters – linking susceptibility to short-term mortality in the MDS

The remaining clinical evidence predicting 30-day outcomes in the MDS is presented in Table 18. Thirty-day mortality differs across comparators via two mechanisms in the MDS. Thirty-day mortality does not differ in the MDS setting if infections are susceptible to existing treatments because patients will be treated with the correct AM, though it does differ if patients have infections resistant to existing options but susceptible to cefiderocol, as their recovery will be more likely if they take cefiderocol. In addition, patients’ mortality risk varies according to the AKI rate associated with the AM used.

As documented in Review questions 4 and 5, several studies have explored the link between whether patients have been administered a treatment to which they are susceptible and their 30-day mortality outcomes in the infection sites of interest for the HVCSs. However, these studies have focused on the ES and none was available relating specifically to the MDS where outcomes are expected to differ substantively. Patients in the MDS who receive an inappropriate drug are much more likely to be MDR than patients receiving inappropriate treatment in the ES and are more likely to be in a critical state that reduces the possibility for further treatment.

This data gap is perhaps unsurprising as multidrug resistance (including to colistin) remains rare and it may, therefore, be difficult to recruit or include sufficient patients in this setting. Given the absence of data to inform this important parameter, a structured expert elicitation exercise was conducted. The methods for the expert elicitation are described in Structured expert elicitation. These estimates were elicited separately for cUTI, HAP and VAP as these infection sites are expected to have quite different mortality rates. Separate estimates were not produced by causative pathogen mechanism. This is because, among those receiving a treatment to which they are susceptible, outcomes are expected to be similar across the pathogen-mechanism combinations relevant to cefiderocol HVCSs. Similarly, among patients receiving multidrug salvage therapy due to multidrug resistance, outcomes are expected to be similar across the pathogen-mechanism combinations relevant to the cefiderocol HVCSs.

Clinical parameters – acute kidney injury risk and subsequent outcomes

Rates of nephrotoxic drug-induced AKI and their implications are assumed to generalise across sites, pathogens and mechanisms in the absence of subgroup-specific data. The evidence from the cefiderocol RCTs provided limited evidence on the safety implications of aminoglycoside/colistin use (see Results) and these data were not considered by our advisors to be generalisable to the highly comorbid patients who tend to acquire carbapenem-resistant infections. These parameters were obtained from existing systematic reviews where possible or, if not available, from UK-specific sources.

Several recent systematic reviews and meta-analyses estimated the pooled cumulative incidence of AKI in patients treated with colistin or polymyxins B,128,129,131,132 and two reported differences in the rates of AKI between colistin or polymyxin B-based therapy and other agents. 131,129 The absolute risk of an AKI and the likelihood that an AKI resulted in irrecoverable kidney damage was derived from Sisay (2021),128 as this study had the most recent searches, included a broad range of study designs and was restricted to studies using the Risk Injury Failure Loss and End-Stage Renal Disease (RIFLE) criteria. The difference between colistin (or polymyxin B, a similar drug from the same class)–based therapy and other agents was obtained from Chien et al. (2020)129 as this review made some attempt to control for confounding. Chien et al. (2020)129 included both RCTs and comparative cohort studies, but excluded studies considered poor quality as assessed by the Newcastle–Ottawa scale (in particular, the authors state that only cohort studies of parallel design with patients with comparable clinical characteristics were included). Alternative sources for these parameters are explored as scenario analyses.

The excess death rate from AKI was derived by comparing in-hospital mortality rates in the UK for individuals who experienced an AKI, as defined by the Acute Kidney Injury Network (AKIN) criteria, and individuals without AKI using the East Kent Hospitals University NHS Foundation Trust (EKHUFT) data set from Kerr et al. (2014). 130 The latter gathers admission records from three inpatient hospitals in the South of England. The analysis of the EKHUFT data set was deemed more appropriate than that obtained using the Hospital Episode Statistics (HES) data set, as EKHUFT includes older and more comorbid patients that are, therefore, more similar to the patient population in this evaluation, and is more likely to include all AKIs than the HES data set. The impact of AKI on mortality was estimated by the authors adjusting for a range of covariates including history of hospital admission, comorbidities and primary diagnosis. We assumed the relative increase in mortality associated with AKI observed in the Kerr et al. analysis applied to the baseline risk of mortality in our HVCSs despite the patients within our HVCSs exhibiting a much higher baseline mortality risk. AKI is more prevalent in patients with poor prognosis and, although Kerr et al. attempted to adjust for these factors, the elevated mortality estimated was considered high by expert advisors. A scenario analysis was, therefore, run whereby the excess mortality associated with AKI was halved from the reported value.

Clinical evidence – linking susceptibility to 30-day outcomes in the empiric setting

The evidence informing the decision tree predicting 30-day outcomes in the ES is presented in Table 19. The mechanisms database described in section Appendix 1 was searched to identify papers providing quantitative estimates of the risk of carrying the pathogen mechanism of interest among patients with specific characteristics. This was supplemented by papers known to the study team. Two searches were conducted. The first was to identify UK-specific studies. This had two concepts; the first was to identify ‘risk’ studies (any field containing any of ‘risk’, ‘prevalence’, ‘incidence’, ‘character*’ or ‘outbreak’), returning 1696 studies. The second concept was for UK studies (abstract contains any of ‘United Kingdom’, ‘Great Britain’, ‘England’, ‘UK’, ‘NHS’, ‘Trust’, ‘London’, ‘*Shire’), returning 119 studies. Combining both concepts provided 61 studies for the first search. The second search was expanded to identify non-UK risk models and returned 51 studies based on their title containing ‘Risk’. No risk models were identified from either search. Indeed, even in the wider population of patients at risk of a carbapenem-resistant infection, there is a paucity of UK data available to estimate the risk of having a carbapenem-resistant infection among patients with relevant risk factors. 133

The probability that a patient entering the ES actually has the suspected pathogen mechanism was obtained from Second-Generation Surveillance System (SGSS) data supplied by PHE as shown in Table 19 (for further discussion see Quantitative extrapolation to expected usage). These data provide the number of tests for a given mechanism of resistance and the proportion of those tests that returned a positive result. For MBL mechanisms, separate tests are run for each specific type of mechanism (i.e. a separate test is run for IMP, VIM, NDM and in the case of P. aeruginosa Dutch imipenemase, DIM). We assume when calculating these figures that, if MBL is suspected, all MBL mechanism tests are run. These data are unlikely solely to reflect the ES HVCSs of focus in the current analysis, for example testing may be conducted due to a suspicion in the lab rather than at the level of the treating clinician (e.g. a lab finding of carbapenem non-susceptibility might trigger a test), some tests may be run following treatment failure or may be run in the ES but at a lower level of suspicion than considered in our HVCSs.

Given these uncertainties in the available data, we also conducted a survey of the mailing list of the BSAC. This survey asked microbiologists and infectious disease specialists how many times they saw patients who would fall into our ES HVCSs of interest, and the proportion of those who actually had the pathogen mechanism of interest. A survey was used in preference to the structured expert elicitation as this parameter was expected to vary according to local epidemiology and history of outbreaks of resistant infections, and it was not considered realistic that the expert elicitation exercise could include enough experts to adequately reflect this geographical heterogeneity. Unfortunately, the response to the survey was low with only nine experts providing usable responses. On average, these experts reported that, of the MBL Enterobacterales or P. aeruginosa HAP/VAP cases seen where there was a high suspicion of the mechanism of interest, 71% of patients would be confirmed as having the mechanism. These values are used in a sensitivity analysis. Given the high level of uncertainty around this parameter, sensitivity analysis results are shown for a wide range of alternative values.

Mortality at the time of assessment for entry to the MDS conditional upon susceptibility status, and 30-day mortality among patients not requiring further treatment, was obtained from Tumbarello (2013). 125 This study was conducted in 110 ICU patients with confirmed P. aeruginosa pneumonia in a hospital in Italy, and compared 30-day mortality in patients who were susceptible to initial empiric treatment and those who were not. Surviving patients who were not susceptible to empiric treatment were switched to definitive therapy, on average approximately 62 hours after symptom onset. Tumbarello (2013)125 was chosen as it reported a relatively high incidence of multidrug-resistant strains in infecting organisms (42/110) compared to the other studies identified in the review and was the only study reporting Kaplan–Meier curves (see Review questions 4 and 5 for details of review). No UK studies were identified.

The probability of requiring further treatment for susceptible patients was taken from the cefiderocol arm of the APEKS-NP study. Within the studies included in the cefiderocol evidence mappings, this was identified as the only study representing a predominantly empirically treated susceptible population of HAP/VAP patients that also reported subsequent treatment rates.

Clinical evidence – long-term mortality

All patients surviving to 30 days face an ongoing mortality risk based on the CARBAR134 and Merrick135 studies. Both studies included UK patients with infections caused by carbapenem-resistant organisms and were, therefore, considered relevant in terms of capturing the highly comorbid nature of patients who acquire these infections. Searches were conducted as described in Appendix 1 but did not identify any further evidence of relevance. A targeted search indicated a lack of data on long-term outcomes in both HAP/VAP and cUTI. It also seemed unlikely that outcomes in all-comer HAP/VAP and cUTI patients would reflect those of MDR patients who tend to have developed MDR infections as a result of multiple contacts with the health system, reflecting a wide range of comorbidities. We therefore chose to focus our review of long-term mortality on patients with resistant infections.

CARBAR134 was used to inform mortality in the base case as it included more geographically diverse patients, had a longer follow-up (2 years compared to 1 year in Merrick) and provided continuous survival estimates over time (i.e. Kaplan–Meier curves). Merrick135 reported all-cause mortality at 1 year of 31% which is similar to the 1-year mortality in CARBAR of 34%.

Kaplan–Meier curves from CARBAR were digitised and a published algorithm136 was used to recover pseudo IPD from the Kaplan–Meier curve for analysis. Parametric survival models were fitted to these data to facilitate extrapolation beyond the observed data. Data from 30 days onwards were used as these were of most relevance to the model. We followed guidance from the NICE Technical Support137 document and fitted a range of parametric survival models: exponential, Weibull, Gompertz, log-logistic, log-normal and generalised gamma. Model fit was assessed according to Akaike’s Information Criteria (AIC), log-cumulative hazard plots, hazard plots and visual assessment of the concordance between model predictions and Kaplan–Meier plots. No specific external data were identified to support validation of long-term predictions, so probabilities of death predicted by each model were compared to general population mortality over 20 years to assess plausibility. A summary of these assessments is provided in Appendix 7, Table 49. Overall, the Weibull, log-logistic and log-normal models were all considered plausible candidates and, in the absence of further evidence, log-normal was selected to offer a middle ground with the Weibull and log-logistic used in scenario analyses. Mortality is restricted so that it remains above that in the general population within the model.

In addition, patients alive with recovered renal function face an elevated risk of death and a risk of developing irreversible renal failure (CKD). 138 Patients alive with irreversible renal failure face the elevated risk of death of CKD patients.

A recent body of evidence, with which our group of experts agreed, suggests that AKI and CKD are closely linked and interconnected, whereby CKD is a risk factor for experiencing subsequent AKI and AKI is a promoter or instigator of CKD. It was, therefore, considered important to capture the fact that AKI is not a ‘self-limited process’ and that patients with recovered renal function post AKI are at risk of adverse renal outcomes and of developing CKD.

In our literature searches to identify evidence of the impact of AKI on the development of CKD and on long-term survival, we looked for studies that would control for the confounding impact of comorbidities as stringently as possible, as we aimed to estimate the causal effect of AKI on subsequent outcomes. The US study by Bucaloiu (2012)138 was selected as it compared outcomes of patients with hospital-associated AKI (with recovered renal function) against a non-AKI patient population matched for a wide range of relevant clinical and demographic characteristics. One thousand six hundred and ten patients with AKI and 3652 without were followed up from 90 days post discharge to approximately 6 years. A limitation of this study is that the propensity score matching process excluded the most comorbid patients due to a lack of sufficiently closely matching controls, and the study excluded patients with impaired kidney function prior to hospitalisation. This evidence was used to inform the increased risk of death and the increased risk of developing CKD in patients with recovered renal function after an AKI. Relevant parameters are shown in Table 20.

There are a number of limitations to the approach taken to reflect the long-term implications of AKI within the model:

1. The CARBAR134 mortality data will have included patients who experienced AKI and, therefore, including additional mortality risk associated with AKI and CKD development is likely to exaggerate mortality risk in the model.

2. The risk of CKD development is likely to be higher than estimated from Bucaloiu138 in the highly comorbid patient group considered within the HVCSs.

3. The hazard ratios on mortality are applied multiplicatively despite the much higher baseline risk of death in the patient population considered within the HVCS.

Scenarios are explored to address each of these assumptions in turn:

1. CARBAR mortality rate is reduced by 10% reflecting an assumed AKI rate of 20% and an assumed excess mortality associated with AKI of 1.48 (95% CI 1.19 to 1.82) based on Bucaloiu (2012). 138

2. Patients in the HVCSs face double risk of CKD compared to patients in Bucaloiu (2012). 138

3. Patients in the HVCSs face an absolute increase in mortality risk observed in Bucaloiu (2012). 138

4. All of the above applied simultaneously.

We did not account for life years accrued within the first 30 days in the model as these were expected to have a marginal effect on the model results.

Health-related quality of life

The HRQoL implications of the infection are not modelled as these are expected to be short- lived and, therefore, are not expected to impact substantively on the model results. However, to quality adjust the life expectancy estimates accurately, we did consider it important to reflect the underlying comorbidities of the patients within the HVCSs. We did not identify any relevant utility data from existing models, most of which assumed that, post infection, patients would return to the HRQoL of the general population. Therefore, we conducted a review of utility studies that provide evidence according to the Charlson Comorbidity Index (CCI). The CCI is a summary score of comorbidity based on 17 included comorbidities. The comorbidities considered in the CCI have been selected and then weighted based on their ability to predict 1-year mortality among hospitalised patients. Importantly, the CCI is reported within the CARBAR study for patients with infections caused by carbapenem-resistant organisms, allowing the utility values presented by CCI score to be reweighted to reflect the CCI scores in a population similar to that included in our HVCSs.

The methods for this review are described in Appendix 1. This identified two studies reporting utilities by CCI in the general population. Both studies were based on large national surveys in France and Germany and estimated the SF-6D based on the SF-36 and SF-12, respectively. The French study was chosen in preference to the German study as the latter controlled for a number of variables likely to be associated with CCI (pain level, sociodemographic variables and health behaviours). Utility values by CCI score are reported in Appendix 7, Table 50; these reflect UK general population valuations. These are weighted by the distribution of CCI scores observed in CARBAR,134 also shown in Appendix 7, Table 50. This produced an overall weighted utility score of 0.66 for the CARBAR population based on their comorbidities; this is intended to reflect their long-term quality of life rather than the immediate impact of infection. This was used to compute a multiplicative reduction in HRQoL associated with comorbidities by comparing the CARBAR population to the general population (assumed to have a CCI score of 0). General population HRQoL was derived from a regression model estimated by Ara et al. (2010)139 using Health Survey for England (HSE) survey data for the years 2003 and 2006 (n = 26,679). This produced a baseline utility value of 0.73 for all patients. This resulted in a utility multiplier of 0.66/0.73 = 0.90. This was then applied to the age and gender-specific EuroQol-5 Dimensions (EQ-5D) utilities of the general UK population based on Ara et al. 139

Patients who have recovered their renal function post AKI are not expected to experience further disutility unless they develop CKD. The HRQoL decrement applied to the CKD patients is computed using pooled estimates from a systematic review and meta-analysis by Wyld et al. (2011). 141 The authors reported decrements of 0.02 (−0.04 to 0.09) for those in CKD pre treatment, and of 0.11 (0.08 to 0.15) for those with CKD in dialysis, where the latter was estimated to represent 2% of the diagnosed CKD population based on UK data. 142 These were applied to the baseline utility value of 0.73 such that the utility of those with CKD pre treatment was 0.71 and the utility of those with CKD in dialysis was 0.62.

Resource use and costs

The model includes costs relating to hospital stay, infection control during hospitalisation, AKI-related costs during hospitalisation, long-term costs associated with CKD and costs relating to use of existing AMs. The purchase price of cefiderocol is not included in the costings as the objective of the evaluation is to inform the payment for cefiderocol. Costs relating to testing (for pathogen, resistance mechanism or AM susceptibility) were not included as, in the HVCS populations, these tests were expected to be conducted to the same degree regardless of the introduction of cefiderocol.

An important cost driver in the model is time spent in hospital. Data on time in hospital for patients according to their treatment pathway and outcomes are presented in Table 21. As for 30-day mortality, we did not identify any studies in the MDS linking treatment susceptibility to duration or type of hospitalisation. This was, therefore, elicited as part of the structured expert elicitation exercise. LoS and the proportion of time spent in ICU or HDU was estimated conditional upon susceptibility for patients with cUTI and HAP/VAP separately. In the base case, all patients in the ES were assumed to spend 5 days in hospital prior to receipt of their microbiology results, the median wait reported in CARBAR. 134

The LoS for patients successfully treated in the ES was estimated from the LoS in patients who are susceptible to treatment in the MDS as estimated from the structured expert elicitation, the time to receiving MDS from CARBAR134 and the relative reduction in the LoS associated with receiving appropriate empiric treatment from Muscedere (2012). 143 The proportion of time spent in the ICU for patients who received a treatment to which they are susceptible and who did not require further treatment was derived from Muscedere (2012). 143 The study was conducted in 350 adult ICU patients with VAP (any pathogen and resistance profile) in Canada who received empiric treatment with meropenem or meropenem + ciprofloxacin. The study reported hospital and ICU LoS in patients who were susceptible to their empiric treatment and those who were not. Muscedere143 was chosen as it was the only study identified in the review in Review questions 4 and 5 that reported LoS conditional upon susceptibility in patients with HAP/VAP. The LoS reported by Muscedere (2012) was skewed. The mean LoS was derived by fitting a log-normal distribution to the reported median and interquartile range. The derived mean LoS in patients who received appropriate and inappropriate treatment (43.1 and 85.7 days, respectively) were used to derive the relative reduction in the LoS associated with receiving appropriate treatment. The derived mean LoS and stay in ICU were used to derive the proportion of hospital stay that was spent in ICU.

The additional hospitalisation costs associated with in-hospital AKI are informed by estimates derived from Kolhe et al. (2014). 144 This study used the NHS costing system’s relative value units that capture cost information associated with several cost items including LoS on wards, drugs, physiotherapy, radiology and medical staff costs.

Unit costs were obtained from standard sources and also considered those suggested in the manufacturer submissions. Where necessary, costs were adjusted to 2019–20 prices using standard sources. 146 The daily cost of cUTIs treated on general medical wards was derived from the weighted average cost of non-elective short stay for kidney or UTIs with/without interventions (LA04H–LA04S). The daily cost of HAP/VAP treated on general medical wards was derived from the cost of non-elective short-stay bronchopneumonia with or without interventions (DZ23H–DZ23N). The daily cost of ICU was assumed to be the weighted average cost of non-specific, general adult critical care (CCU01) with zero to six organs supported (XC01Z–XC07Z), assuming that ventilation cost is reflected in the organ support costs. The daily cost of HDU was assumed to be the weighted average cost of medical adult patients in critical care (CCU03) with zero to six organs supported (XC01Z–XC07Z). Weighting was based on the overall volume of each type of organ support reported for the NHS. The daily cost of isolation was derived from Knight et al. (2018),145 and included the cost of gloves, aprons and infectious waste stream. It was assumed that all patients would be subject to isolation measures as they are either highly suspected of having or confirmed to have an MDR infection. One-off costs of stock disposal are not included as these are assumed to apply equally to all patients.

Following discharge, patients’ long-term costs are determined by their health state. Patients alive without a history of AKI, or with recovered renal function, experience no further costs. Patients with irreversible renal failure (i.e. CKD) face a weighted average cost that reflects the CKD-severity distribution in England and requirement for dialysis. Kerr et al. (2012)142 estimated the annual per patient NHS expenditure on CKD direct care, dialysis and transplants. The mean annual cost of direct CKD care per patient not on dialysis (i.e. antihypertensive drugs, primary care tests and consultations, nephrology consultations and cost due to excess incidence of cardiovascular events) was estimated at £278, while the annual cost of CKD-related care for a patient on dialysis was estimated at £31,933. These costs are similar to those proposed by Shionogi who estimate that patients with nephrotoxicity not experiencing dialysis face a cost of £173 in year 1 and £89 from year 2 onwards (based on assumed resource use), and £24,000–29,000 for those receiving dialysis (source unclear). As for quality of life, 2% of the diagnosed CKD population were estimated to be receiving dialysis based on UK data. 142 The clinical advisors to EEPRU for this project indicated that this may be an overestimate, but use of a lower value is unlikely to substantially change the results of the modelling. Our clinical advisors expected that, in the highly comorbid group considered within the HVCSs, transplant would be rare; the costs of transplant were not, therefore, included in the CKD cost estimates. This results in a weighted average cost of CKD of £911 per annum in 2019–20 prices.

We did not include differential rates of discharge to long-term care facilities in the base-case analysis as no evidence was found comparing UK usage of care among those with and without AKI that adjusted for differences between patients with and without AKI. US data suggest that AKI is associated with an elevated risk of discharge to long-term care even with adjustment for other predictive factors. Liangos (2006)147 found that 8.9% of patients without AKI will be discharged to long-term care, and this is elevated to 17.8% in those with an AKI (reflecting an adjusted OR of 2.2, 95% CI 2.1 to 2.2). We combine this with information on the costs of long-term care and model a scenario based on this. We use a weekly cost of £1049 (average of private sector nursing home, and local-authority own-provision residential care for older people146) and apply this for the lifetime of the patient. This is likely to be an overestimate as some patients may be discharged from long-term care and the full cost of this care may not fall on the NHS/Personal Social Services (PSS) budgets.

We did not include the cost of end-of-life or palliative care as this was considered unlikely to substantially influence model outcomes.

Drug acquisition costs were based on the cost of the daily dose derived from published sources,148,149 the daily doses reported by WHO Collaboration Centre for Drug Statistics Methodology150 and the treatment duration derived from published literature. 149,151–153 When more than one formulation or pack size was available, we based costs on the largest pack size and IV formulations. When the treatment duration was provided as a range, we used the longest duration, to reflect the high severity of infections. When more than one AM was available for treatment in a particular setting (e.g. colistin or aminoglycosides for the treatment of HAP/VAP in MDS), the most expensive treatment was chosen to reflect that often combination or higher doses of therapy may be used. In the ES, patients who require a treatment switch following availability of their susceptibility results are assumed to receive 5 days of treatment, whereas those who do not require a treatment switch receive the full course. The unit cost of all comparators is shown in Appendix 7, Drug acquisition costs, Table 51. The drug acquisition costs used in the model are summarised in Table 22.

Predicting the future sizes of the high-value clinical scenarios

The objective of this analysis was to statistically model changes in the number of patients within the HVCS over time, to inform a quantitative forecast of the number of patients presenting in the HVCS over the next 20 years.

Data on the number of infections over time for the mechanisms of interest were provided as a time series by PHE. Two populations were included:

• Enterobacterales with a MBL mechanism

• Pseudomonas aeruginosa with a MBL mechanism.

Data were supplied for invasive infections which are predominantly infections where the specimen sample relates to a BSI or cerebrospinal fluid infection. It was assumed that, for each pathogen mechanism of interest, the trends in population size for invasive infections generalise to the HVCSs. This was considered reasonable by the clinical advisors to the project. The small number of invasive infections made it challenging to reliably identify if there was a trend in the growth of the HVCSs. As such, this analysis is supplemented by a secondary analysis which looks at trends in the number of screening isolates. These isolates are from screening specimen sites. Screening samples were broadly categorised as samples from swabs, wounds and the lower gastrointestinal tract. It includes potential infections as well as isolates from people who do not have infections but may be colonised by a MDR pathogen. These screening isolates were only used to confirm or refute the potential presence of a trend rather than inform the growth estimates as they may be influenced by screening policy changes over time which may not feed through to changes in identified infections. Data on both invasive infections and screening isolates were obtained from the Antimicrobial Resistanceand Healthcare-Associated Infections (AMRHAI) national reference laboratory. These data were provided by PHE as monthly counts and are available from 2004 (P. aeruginosa: from 2003) to April 2021. During 2018, guidance on which samples should be sent to AMRHAI changed, and charges were introduced. This led to a gradual ‘artificial’ decrease in referrals. Further detail on the nature of this data set is provided in Report Supplementary Material 1.

Due to small numbers, data on invasive infections were aggregated to quarterly for analyses and restricted to October 2012 onwards. The last observations used were for March 2018 (inclusive), as after this point the observed numbers decreased. For screening isolates numbers were larger, so monthly data were used. For these, the first observation was set to be the first time point for which there were no future months with zero counts (April 2013 for Enterobacterales and November 2014 for P. aeruginosa).

Time-series (state-space exponential smoothing) models were used to forecast the isolate data. For the invasive isolates, the use of other time-series models was also considered. Further details on the models considered and the justification and implementation of the state-space models is provided in Appendix 7, Further details on modelling direct population net health effects in high-value clinical scenario. Three state-space models were considered. These varied with regard to the assumptions made about any long-term trends in the growth of the HCVS:

• no growth (no trend)

• growth in the short term that in the long term changes to no growth (a ‘damped trend’ model; the degree of dampening is estimated from the data and influences how quickly the growth tends to zero)

• persistent growth (trend that is not damped).

Within-sample goodness-of-fit statistics for the three models and the two data sets are provided in Appendix 7, Goodness of fit of forecast models, Table 52. Estimates of population growth are provided in Figure 13 (estimates are not shown for P. aeruginosa as the best-fitting model was always one without a trend). The two isolate data sets are very different with regard to absolute numbers. Figure 13 shows the change in population size over time for both the data set of invasive infections and the screening isolates.

FIGURE 13.

Change in population size over time (top pane = invasive isolates, bottom = screening isolates). MBL, number of isolates with Enterobacterales metallo-beta-lactamases.

For the invasive isolates, for both the Enterobacterales and P. aeruginosa populations, the best-fitting model (based on within-sample fit) is one without a trend. Of the two models that include a trend, the trend model provided the best fit for the Enterobacterales population, while the damped trend model provided the best fit for the P. aeruginosa populations. Visually, there was no evidence of a trend for the P. aeruginosa, while for the Enterobacterales, there was a potential trend.

For the screening isolates, the best model for the P. aeruginosa was again one without a trend, while for the Enterobacterales, it was a model with a non-damped trend. However, for the screening isolates, it is unclear if long-term increases in Enterobacterales reflect genuine increases or the results of increased testing. It is also unclear if any genuine increases would persist into the future. For Enterobacterales, the largest relative change was for the screening isolates with a non-damped trend and the smallest (non-constant) change was for the invasive isolates with a damped trend model. As long-term estimates were very sensitive to the choice of model, and there were little statistical grounds to choose between the two models, both the damped trend and trend models for invasive isolates were considered within the decision-analytic modelling for MBL Enterobacterales. It was assumed that there was no trend in the future size of P. aeruginosa.

Details on how the estimates of future change in the HVCSs were used in the economic model are provided in Appendix 7, Further details on modelling direct population net health effects in high-value clinical scenario.

Predicting future rates of resistance for current practice

The objective of this analysis was to characterise historical changes in resistance to existing AMs among patients with MBL Enterobacterales and MBL P. aeruginosa to inform a quantitative forecast of how resistance might change in the future.

This analysis used time-series data provided by PHE, obtained from the same evidence sources as described in the previous subsection (i.e. the AMRHAI national reference unit). Analyses were restricted to comparators used in the economic model. Resulting data were available for:

Enterobacterales with MBL:

• aminoglycosides (gentamicin, amikacin, tobramycin)

• aztreonam

• colistin

• tigecycline

Pseudomonas aeruginosa with MBL

• colistin.

For AM classes with evidence from multiple AMs, the most resistant result was retained in the supplied data. It is not expected that retaining the least resistant result would have a noticeable impact on estimates of resistance over time. As already described, isolates that were reported as ‘intermediate’ resistant were assumed to represent resistant isolates for the purpose of this analysis. Hence, any tested isolate was either categorised as ‘susceptible’ or ‘resistant’ for this analysis. Combining ‘intermediate’ and ‘resistant’ categories was based on advice from clinical advisors. It is, however, noted that current EUCAST guidance is to combine ‘intermediate’ and ‘susceptible’ when only two categories are used. 159 The methods used to generate forecasts are broadly the same as those considered in the previous section, and are discussed in more detail in Appendix 7, Further details on modelling direct population net health effects in high-value clinical scenario. Due to the sparsity of the available evidence, trends in the resistance (or susceptibility) to comparator AMs were not incorporated into the decision-analytic model (for further details, see Appendix 7, Further details on modelling direct population net health effects in high-value clinical scenario, Table 53).

Predicting future resistance trajectories for cefiderocol

The time-series data on susceptibility provided by PHE (and detailed in the previous subsection) do not include cefiderocol. In the absence of evidence for the drug of interest, trajectories for the development of resistance were identified for other AMs. Due to a lack of evidence, the searching to identify these trajectories were not restricted to the HVCS populations. Evidence linking these resistance trajectories to quantity of AM use was also sought. This was to enable a link with the economic modelling; differences in levels of cefiderocol use (e.g. in response to an increase in the HVCS populations over time or when comparing use in the ES with use in the MDS) are expected to be associated with differences in the rate at which resistance is gained. To emphasise that the evidence used is for a different drug and patient population, it is referred to as ‘external AM use-resistance data’. Including an association between cefiderocol use and the development of resistance means that scenarios with increased levels of cefiderocol use will not automatically result in better long-term outcomes. There is a large body of literature demonstrating that AM use is associated with subsequent resistance. 160–162 As there are no historic data on how resistance develops for cefiderocol (in either the PHE data described in the previous section or identified in the susceptibility searches of Clinical evidence), any resistance trajectories used in the economic modelling will be subject to considerable uncertainty. This is particularly pertinent as the relationship between AM use and resistance has been shown to vary by both type of AM and geographical setting, and in some situations, there is no apparent relationship. 161,163,164 In the company submission for cefiderocol, it was contended that the development of resistance may be lower for cefiderocol than historically observed for other AMs (see section 2.3.5 of the company submission82). Conversely, as an AM with broad coverage, the development of resistance to cefiderocol may be similar to other broad-spectrum AMs (which is typically quicker than the development of resistance for narrow-spectrum AMs). This emphasises the importance of identifying the robustness of estimates of population-level INHE to different potential resistance trajectories.

Two approaches were used to identify external AM use-resistance data that may inform the use-resistance association. First, the entire database of studies that was used during the reviewing process (for both CAZ-AVI and cefiderocol) was searched. Studies were filtered to include those which included ‘use’, ‘usage’, ‘volume’ or ‘consumption’, and these were searched for any relevant evidence. In addition, to identify any English studies (which may use evidence from PHE, or the online portal ‘fingertips’), a Web of Science search was conducted with the terms ‘(AM* OR antibiotic* OR resistan*) AND (fingertip OR “Public health England”)’. These searches were complemented by any studies that were identified via other reviewing activities or already known to the study team. As a result, three studies were identified that, while not using data in the public domain, provided information on a use-resistance relationship. 165–167 Details of these studies are provided in Appendix 7, Further details on modelling direct population net health effects in high-value clinical scenario, Table 54. These existing studies informed the de novo analyses reported here by suggesting that autoregressive integrated moving average (ARIMA) models would be suitable time-series models for capturing use-resistance associations, with a lag of 1 year between use and resistance when using annual data.

In addition, several studies used publicly available surveillance data. 163,168–172 These data were reanalysed for this project to identify potentially useful associations. For this project, there were two types of data of interest:

• English data on AM use and AM resistance, from the ‘AMR local indicators profile’. 173

• European data on AM use and AM resistance from the European AM Resistance Surveillance Network (EARS-Net) and European Surveillance of AM Consumption Network (ESAC-Net), respectively. 174,175 These are available as annual data.

Further details on these evidence sources are provided in Report Supplementary Material 3. For cefiderocol, increases in resistance will be from a low starting point. Observed trajectories for external evidence which also showed an increase from a low starting point were only identified for the European data and so this was used in subsequent analyses.

Thirty countries from the European Union contribute data to EARS-Net on AM resistance for up to eight pathogens. 176 These data were further filtered based on the following criteria:

• Pathogen is included in the HVCSs (E. coli as an Enterobacterales and P. aeruginosa).

• Data were available for both AM use and AM resistance (cephalosporins of all types, and carbapenems).

• Countries with at least 5000 isolates were tested, baseline resistance (average over the first 3 years of available data) was < 3% (Enterobacterales) or < 15% (P. aeruginosa), with at least 10 years of observations for carbapenems and 15 years of observations for cephalosporins (these did not have to be consecutive).

This resulted in the following 23 pathogen–drug–country combinations:

• Pseudomonas aeruginosa, carbapenems: Finland, France, Ireland, Netherlands, Norway, Slovenia, Sweden

• E. coli, carbapenems: France, Greece, Netherlands, Norway

• E. coli, cephalosporins: Bulgaria, Croatia, Estonia, Finland, France, Greece, Ireland, Luxembourg, Malta, Norway, Slovenia, Sweden.

For these countries, ARIMA models were used to estimate the impact of increasing AM use (defined daily doses per 1000 inhabitants per day) in a given year on resistance to that AM in the following year. The results of these analyses are shown in Appendix 7, Further details on modelling direct population net health effects in high-value clinical scenario, Table 55. Of the 23 combinations considered:

• Just under half provided a significant association (12/23; P. aeruginosa = 4/7, E. coli = 2/4 for carbapenems and 6/12 for cephalosporins).

• Of the 12 significant associations, 7 were positive associations (increasing use led to an increase in resistance), while 5 were negative (decreasing use led to an increase in resistance). Four of the negative associations were for E. coli cephalosporins, the remaining one was for P. aeruginosa.

Hence, this analysis resulted in up to seven significant positive associations that could be used to link increases in AM use to AM resistance in the economic model. Increases in AM use are driven by increases in the eligible population over time.

Projections of expected usage for cefiderocol from Extrapolation from high-value clinical scenario to expected usage were linked to these estimates of the relationship between usage and resistance to predict emergence of resistance to cefiderocol over time. Even under more extreme usage predictions and the strongest associations between usage and resistance emergence, this predicted small absolute increases in resistance up to 0.04% and 0.16% over 20 years for the MBL Enterobacterales and P. aeruginosa populations, respectively (see Report Supplementary Material 3 for more details). This may represent an underestimate of the potential for resistance emergence for two reasons. Firstly, the spread of MDR infections is influenced by international travel and the ‘importation’ of MDR pathogens. Resistance emergence may, therefore, be influenced by cefiderocol usage outside the UK which is not accounted for in these projections.

Secondly, the relationships between usage and resistance characterised in the available data reflect all tested isolates in the community and hospital settings. Resistance emergence may be much higher within the HVCSs where usage will be concentrated. For this reason, EEPRU has conducted a range of scenario analyses to characterise the potential emergence of resistance to cefiderocol. These were informed by considering the absolute increases in resistance for the drug–pathogen combinations and countries discussed above, where there was a statistically significant increase (see Report Supplementary Material 3 for more details). The highest absolute increase in resistance (an annual absolute increase of 1.65%, leading to a projected 20-year increase of 33%) was used to bound these analyses. The second largest increase was 0.95% per year (19% over 20 years). Based on these considerations, EEPRU ran analyses with resistance emergence reaching 1%, 5%, 10% and 30% at 20 years. It is noted that the upper scenario may be very extreme.

Of note, this analysis was focused on data sets which demonstrated an increase in resistance over time. Hence, any significant associations between AM use and decreasing resistance were not explored. As an alternative to an ARIMA model, a dynamic differential equations model was also developed. This was designed to incorporate AM use and resistance, as well as the spontaneous loss or gain of resistance over time as well as the impact of deaths. Details of this model are provided in Report Supplementary Material 3; when evaluated in a simulation study, it was shown to provide biased parameter estimates. This was potentially due to the non-identifiability of the model (due to the number of potential AM drivers considered), so this model was not considered further.

As a face-validity check of the estimates of AM use employed in the model, these were compared to hospital inpatient drug use as reported in the 2019/20 ESPAUR report. 6 This provided an estimate of 2.4 DDD/1000 inhabitants for all AMs used in an inpatient setting. Drug use during the first year of the economic model for both CAZ-AVI and cefiderocol (combining results from both evaluations) for the sites cUTI, IAIs, HAP/VAP and BSIs (all four in Enterobacterales, P. aeruginosa, and Stenotrophomonas) was estimated to be 0.00018 DDD/1000 inhabitants, hence representing 0.01% of all hospital inpatient AM use. This estimate, as an upper bound on the potential use of both cefiderocol and CAZ-AVI, was felt by the modelling team to have face validity.

Predicting the impact of reduced drug use on resistance

Introducing cefiderocol (compared with the situation when it is not available) may lead to a reduced use of comparator AMs. As the economic model includes an association between increased use of cefiderocol and increased resistance (as described in the previous subsection), then intuitively a decrease in AM use would be expected to lead to a decrease in resistance. However, AM use in the population of interest is only one of a multitude of potential drivers for increases in AM resistance. Other potential drivers include the number of invasive procedures, AM use in other countries, environmental factors and AM use in animals. 177,178 The existing evidence on the effect of reduced AM use on AM resistance is mixed,179 with findings including no decrease, a decrease and even an increase in AM resistance. 180–182 Hence, while the introduction of a new AM is expected to lead to an increase in resistance over time, reducing AM use has less predictable effects on resistance. Due to the heterogeneity in the existing literature and the lack of evidence for the population and AMs of interest, it was assumed that reductions in the use of existing AMs did not lead to reductions in resistance over time.

Areas of expected usage

Quantitative extrapolation to expected usage

Enablement value

Improved treatment of pre- and postoperative MDR infections is included within our HVCSs and expected usage projections. There is some uncertainty as to whether the full benefits of treatment of preoperative infections are reflected within our analysis. We assume that all patients who are alive at 30 days experience the same survival. If, however, the speed of resolution of an infection influences whether a procedure or treatment can go ahead, then it is possible that 30-day survival is longer for patients whose infection resolves more quickly as they may be more likely to receive procedures. The magnitude of this effect is uncertain due to uncertainties about the number of patients who experience infections preoperatively, the impact of infection duration on the likelihood that operations will go ahead, and the implications of operations not going ahead (which will depend on the type of procedure and whether the procedure is not conducted at all or delayed).

Enablement value may be also realised if the risk of MDR infection and clinicians’ ability to treat an MDR infection influences a decision about whether to bring a patient in for a procedure. An example of this scenario was provided by our clinical advisory group, whereby if an MDR infection is known to be circulating in a haematology unit, certain patients may not receive planned procedures or treatments. This is particularly likely to apply for patients in whom existing antibiotics for MDR infections are not an option. Here, the specific example of myeloma patients was highlighted as such patients are predisposed to renal impairment which rules out key effective treatments for MDR infections such as colistin. There is uncertainty with respect to the number of patients who would be affected as this would depend on both the number of patients whose treatment would be impacted by an outbreak and the frequency of outbreaks in key units such as haematology. There is also uncertainty about the consequences for patients not receiving planned therapy, as this will depend on the nature of the procedure/treatment and whether therapy is not received at all or delayed. These effects are not captured within the EEPRU modelling or any quantitative assessments submitted by the manufacturer.

A related way in which enablement value may be realised is if the availability of effective treatments for MDR infections allows wards to be kept open in the face of outbreaks. EEPRU considers it unlikely that cefiderocol would have this effect as most patients with drug-resistant infections do have alternative (albeit more toxic) treatment options – namely colistin. These effects are not captured within the EEPRU modelling or any quantitative assessments submitted by the manufacturer.

A final way in which enablement value may be realised is by use of cefiderocol freeing up healthcare resources. For example, use of cefiderocol may reduce time in hospital (alleviating pressure on beds), including time in the ICU/HDU. This may be particularly important where patients with MDR infections consume additional resources and staff time due to the need for additional infection control procedures including isolation measures. Any freed-up resources can then be repurposed to provide care for other patients within the hospital. To the extent possible, the impact of cefiderocol on resource use has been captured in the EEPRU modelling. When calculating the INHEs of cefiderocol, we have translated cost savings to health benefits using standard measures of health opportunity cost (which allow monetary savings in healthcare resources to be translated to health gains across the NHS).

Diversity value

Our clinical advisors indicated that, within the HVCSs, diverse prescribing strategies (e.g. randomly allocating patients with similar clinical indications to different treatments) were unlikely to be appropriate given the lack of safe and effective alternative treatments. They were not supportive of the use of cefiderocol in broader populations as part of a diverse prescribing strategy due to the desire to reduce emergence of resistance to cefiderocol and concerns that the evidence for diverse prescribing was uncertain. This is in contrast to the views of the manufacturer who emphasised the potential role of cefiderocol in OXA and KPC populations where, due to the availability of other effective treatment options, it could be used as part of a diverse prescribing strategy. Diverse prescribing strategies were not, therefore, included in our quantitative assessments of population-level INHEs.

There is uncertainty about how reduced use of existing agents (e.g. colistin) due to availability of the cefiderocol will contribute to the emergence of resistance to these drugs. Due to these uncertainties, this was not reflected in the EEPRU modelling. If the reduced use of existing agents reduces resistance to existing drugs within areas of expected usage for cefiderocol, this will reduce the INHEs associated with cefiderocol. However, if resistance to existing agents reduces outside areas of expected usage for cefiderocol, this will increase the INHEs associated with cefiderocol. Given the potential for these countervailing effects, and the wide range of factors driving resistance to existing drugs, this is not expected to have a large impact on INHEs.

Insurance value

The scenarios modelled can be considered to reflect this form of insurance value as they involve heavily restricting usage to preserve long-term effectiveness. It is generally agreed that the value of cefiderocol will depend on the trajectory of emergence of MDR infections over time. Within the HVCSs, we have used statistical forecasting methods and explored uncertainty around these to understand the possibility that cefiderocol results in the avoidance of significant/catastrophic health losses. This is presented as distributions of population-level INHEs to inform the committee’s deliberations about whether avoidance of these extreme events should be considered differentially to other forms of health losses.

There is uncertainty around whether these distributions adequately reflect the uncertainty around high-consequence/low-probability outcomes. There are several reasons why this may not fully quantify longer-term benefits:

1. We may see higher levels of resistance to existing drugs within the areas of expected usage over time.

2. We may see MDR pathogens, against which the new drugs are effective, emerge that are currently rare or even unknown.

3. We may see pathogens that are currently treatable with existing therapies (and are not, therefore, included in the areas of expected usage) become resistant.

If one or more of these factors emerge, they would be expected to occur in the long term and quite possibly beyond 20 years.

Conducting quantitative modelling of these effects is unavoidably highly speculative. However, on request from the NICE committee, EEPRU developed an additional scenario to explore the magnitude of these effects. The scenario aimed to assess the effect on INHEs of cefiderocol in case of emergence of multiresistant pathogens against which cefiderocol is the only effective treatment and, in that product’s absence, clinicians would be forced to use multidrug salvage therapy. A simple Excel tool was developed to explore the potential magnitude of these effects informed by parameters obtained from the NICE Committee. This showed an additional benefit of 0.8–58.4 QALYs, which is low relative to other expected benefits of cefiderocol, as estimated by the main analysis (896–3559 QALYs). This reflects the low likelihood of the emergence of a highly resistant strain and the low number of individuals affected by the strain in the medium term. Further details of the methods and results of this additional analysis are presented in Appendix 8, Additional analysis to reflect insurance value. The patient-level INHE used within this analysis and other relevant parameters are shown in Appendix 8, Additional analysis to reflect insurance value, Tables 63 and 64, and the population-level results are shown in Appendix 8, Additional analysis to reflect insurance value, Table 65.

Transmission value

Our clinical advisors indicated that the direction of effect of introduction of cefiderocol on transmission was uncertain, but that, overall, the magnitude of effect was expected to be small. This reflects the fact that introducing a new effective drug for the treatment of MDR infections has a number of countervailing effects. If the drugs reduce time in hospital, this is expected to reduce transmission. However, among MDR patients with poor prognosis, more effective treatments may, feasibly, increase the time spent in hospital by reducing mortality. In addition, where the use of the new drugs reduces mortality, this will increase the number of people returning colonised to the community, as cefiderocol was considered unlikely to eradicate colonisation by the clinical advisors to this project. This may contribute to increased transmission in the community or via further hospitalisations in this highly comorbid population.

The key drivers of transmission of MBL Enterobacterales and MBL P. aeruginosa are broad and driven by transmission in populations beyond the HVCSs (e.g. colonised individuals in the community and in the hospital, and importation of drug resistance from abroad), making this a challenging area to model. Given the views of our clinical advisors that this would not be a key driver of population-level INHE and these modelling challenges, we did not attempt to quantify transmission value using transmission modelling.

To support the committee in its decision-making we do, however, provide a summary of the impact of each drug on time in hospital and time alive post discharge. Briefly, cefiderocol led to a short reduction in the hospital LoS (< 1 day on average), and increased the length of life by 7–84 days (0.02–0.23 years) depending on the subgroup. We note that time post discharge is likely to include further periods spent in hospital given the patient population though we did not quantify these.

A number of advisors discussed the substantial impact of outbreaks of MDR infections in terms of disrupting healthcare provision and incurring large costs due to the need for more extensive infection control measures. However, no evidence was provided that cefiderocol would substantially impact on the likelihood of an outbreak or its spread. The possibility of outbreaks leading to large numbers of cases and the additional potential value of cefiderocol as a treatment in this scenario is discussed under insurance value.

Spectrum value

Our clinical advisors and other stakeholders did not consider spectrum value to be a significant consideration for cefiderocol which has a broad spectrum of activity. Therefore, this was not considered in our quantitative assessments of population-level INHE. The manufacturer, however, notes that:

cefiderocol is also expected to have low impact on the gut microbiota and cause minimal collateral damage to patients due to its narrow aerobic Gram-negative spectrum and predominantly renal clearance with negligible excretion into faeces. Therefore, it would likely not contribute to further disruption of the protective gut microflora and would not add to selective pressure for persistence of MDR and carbapenem-resistant pathogens in the GI tract or contribute to CDI. 81