Relevant population(s)

The baseline population and prevalence of CKD in hospital for the model were obtained from a Grampian population cohort (described in Model structure: initial decision tree phase). The model base-case analysis is therefore based on a mixed cohort of CKD and non-CKD patients, with an average age of 63 years, and a 54.3% female population.

Diagnostic biomarkers evaluated

The model aims to assess the cost-effectiveness of the NephroCheck test, the ARCHITECT urine NGAL assay and the BioPorto urine and plasma NGAL assays in combination with standard clinical assessment, compared with standard clinical assessment alone (including serum creatinine and urine output), for evaluating critically ill people at risk of developing AKI who are being assessed for possible critical care admission.

Model structure: initial decision tree phase

The systematic review did not identify any randomised trials providing causal evidence for the effect of biomarker-guided care versus standard monitoring (serum creatinine)-guided care on patient-relevant outcomes such as peak AKI severity, admission to ICU, need for RRT, CKD or mortality.

In the absence of such data, the initial decision tree phase of the model used a linked-evidence approach to capture the potential impact of diagnostic test accuracy (sensitivity and specificity) on the probability of averting AKI or reducing its severity through earlier adoption of a KDIGO care bundle triggered by a positive biomarker test result. The model then captured possible effects on changes in health outcomes through associative links between AKI severity and the relevant outcomes [need for ICU care, length of stay (LOS), 90-day mortality, and development of CKD].

These associative links have been built up in the decision tree by reanalysis of observational data from Grampian.¹⁰² The data set includes 17,630 adult patients admitted to hospital in Grampian in 2003 and is the complete population of all patients who had an abnormal kidney function blood test on hospital admission, including all patients who developed AKI. The study methodology is described in detail by Sawhney et al.,¹⁰³ but the data derived from the data set used to populate the model are unpublished. These observational, population-level data were used to define the starting age, sex and underlying proportion of prevalent CKD cases in the modelled cohort. The data were also used to populate the model with respect to the distribution of peak AKI severity, as well as LOS in hospital, probability of admission to ICU and 90-day mortality parameters (by KDIGO AKI stage) for the decision tree phase of the model.

In the decision tree, patients who are critically ill in hospital, are at risk of developing AKI and are having their kidney function monitored are divided into two cohorts, those with AKI and those without AKI, depending on the underlying prevalence.^102,103 The underlying prevalence of AKI was calculated directly from a more recent version of the Grampian data set,¹⁰² describing all hospital admissions with at least one overnight stay in 2012 (for patients having their kidney function monitored). The base-case prevalence of AKI generated from these data was 9.2%, sampled probabilistically from a beta distribution in the model based on count data. A sensitivity analysis uses prevalence data directly from the systematic review studies used to generate the diagnostic test accuracy parameters.

Acute kidney injury is defined in the model as having, or being destined to develop, AKI while in hospital and is classified based on the peak severity of AKI. There is an assumption in the model that it is possible to avert AKI with early biomarker-guided treatment in people who would otherwise develop it under standard care. However, it should be noted that, in some circumstances, it may not be possible to avoid AKI by earlier detection, as AKI may not always be modifiable.³ The probability of averting AKI is zero in the standard-care arm. AKI is split into four KDIGO-defined stages (stages 1–3), with stage 3 split by the proportion of patients receiving RRT or not. The initial phase of the model describes the associations between peak AKI classification and probability of admission to ICU, LOS in ICU, LOS in hospital and 90-day mortality. These associative effects are all derived from the Grampian population cohort described previously. At the end of the 90 days, costs and QALY payoffs are assigned based on the decision tree pathway followed, before surviving patients enter the Markov cohort model.

The standard-care cohort is assumed to be perfectly identified as having or not having AKI, based on a combination of serum creatinine levels, other diagnostic workup and clinical expert opinion, which represents clinical practice. The hypothesised advantage of the biomarkers is that they may help to detect AKI earlier, but will not detect additional cases of AKI not detected by current practice. Figure 20 provides an illustration of the initial decision tree pathways for the standard-care arm of the model.

FIGURE 20

Simplified decision tree structure up to 90 days for the standard-care (serum creatinine) arm of the model. Note that the AKI 3 pathway in the model is replicated for the proportion of the cohort receiving acute RRT and those not receiving acute RRT. (more...)

Participants in the intervention (test) arms of the model are similarly split into those with and without AKI, according to the same prevalence data, but all participants receive additional testing. It is assumed that the background diagnostic workup is similar for all arms of the model (i.e. all patients will continue to have their serum creatinine and urine output monitored). As the diagnostic accuracy test data are primarily based on single use of the test, it is assumed, in the base-case model, that each test will be administered only once. It is assumed that the test is administered as soon as possible after a patient has been determined to be at risk of AKI to enable early detection and preventative measures to be implemented. A sensitivity analysis explores the impact of more frequent multiple-use tests on the results.

The diagnostic accuracy of the candidate tests in addition to serum creatinine, compared with serum creatinine alone, was obtained from the results of the systematic review and meta-analysis described in Chapter 3. Table 15 describes the diagnostic accuracy parameters, namely sensitivity and specificity, used in the modelling. All diagnostic data are incorporated probabilistically in the model, accounting for the joint uncertainty in sensitivity and specificity for each biomarker test. The logit of the sensitivity/specificity for each of the biomarker tests was derived from the meta-analysis of diagnostic accuracy studies. The model specified the correlation between sensitivity and specificity parameters (on the logit scale). These parameters were converted to Cholesky decomposition matrices, with the decomposed data referenced by multinormal distributions, sampling from the mean and standard error (on the logit scale). The probabilistic draws were back-transformed from the logit scale for application in the model. It should be noted that diagnostic accuracy data obtained from the meta-analyses are based on heterogeneous studies with different thresholds; this is particularly true for the NGAL assays. Therefore, the results of the economic model, particularly for comparisons between different NGAL assays, should be interpreted cautiously. Further details have been provided in Chapter 3.

TABLE 15

Sensitivity and specificity data used in the model

For each biomarker test group, the proportions of true AKI cases that are true positive and false negative are determined by test sensitivity, whereas the proportion of non-AKI cases that are true negatives or false positives are determined by test specificity.

Based on the EAG’s own clinical expert opinion (Simon Sawhney and Callum Kaye, University of Aberdeen, 2019, personal communication), it is assumed in the base case that patients testing negative would not have any adaptions made to their care pathway. This is because it would be unlikely that care would be de-escalated based solely on a negative NephroCheck or NGAL result, as the conservative practitioner would wait to ensure that there was no rise in serum creatinine before concluding that no AKI was present and stepping down the patient’s care.

The model assumes that all patients will receive the KDIGO care bundle once they are defined as AKI positive using current standard practice methods (i.e. monitoring serum creatinine and urine output), regardless of their NephroCheck or NGAL test result. The potential to benefit from use of the biomarkers, therefore, lies in the early adoption of a preventative care bundle. For patients testing positive, the model includes the functionality to reflect uncertainty in clinical decision-making, that is the probability that a positive test would be acted on. This parameter is assumed to take a value of 100%, in accordance with best practice guidance whereby positive biomarker tests should have a preventative KDIGO care bundle implemented, with the associated costs. Although all positive test results will trigger the KDIGO bundle, only those patients whose tests are true positive will accrue any potential benefits of having their AKI averted or having reduced severity (i.e. peak KDIGO stage) AKI. For exploratory scenarios in which a test might not be acted on in practice, the cohort would follow the standard-care pathways according to whether or not they had AKI, as measured using current clinical practice.

There is limited direct evidence to describe the impact of the use of the AKI biomarkers on important health outcomes (such as need for ICU care, length of hospital stay, risk of 90-day mortality or development of new/progression of existing CKD). Therefore, a linked-evidence approach was required, whereby we relied on observational associations to infer how prevention or mitigation of AKI may affect changes in health outcomes. The associative effects are the benefits of averting or mitigating AKI that lead to better health outcomes (i.e. need for ICU care, CKD and mortality).

These associations necessitate causal assumptions, but, although a causal link between AKI and poor outcomes is plausible, the extent of this causal relationship is uncertain and controversial. It cannot necessarily be assumed that, by averting or changing the severity of AKI, a patient would have the exact same risks (associative effects from the Grampian observational data described previously) of ICU and mortality as a patient who was never going to develop AKI in the first place.

As the true causal relationship between AKI and health outcomes is unknown, the model includes the functionality to apply none, all or a proportion of the relative risk (RR) of health outcomes such as need for ICU care, mortality and CKD (AKI vs. none) to the AKI-averted proportion of the cohort. This is achieved while maintaining the observational associations in the standard-care arm of the model.

The base-case analysis assumes that there are no adverse health consequences of a false-positive test on either NephroCheck or NGAL. Clinical expert opinion (Simon Sawhney and Callum Kaye, personal communication) indicates that there may be a risk to a patient’s health of inappropriate fluid resuscitation; delay of access to appropriate imaging because of concerns regarding contrast exposure; or removal of the most effective, but potentially nephrotoxic, treatments for a critically ill patient. However, the magnitude of this negative effect is difficult to quantify. Therefore, a sensitivity analysis explores scenarios in which an additional mortality risk is added for false-positive tests.

In summary, the early-stage (up to 90 days) costs and outcomes depend on (1) the diagnostic accuracy of the test, (2) clinical decision-making in the presence of positive or negative test results, (3) the initiation of a KDIGO care bundle to avert AKI and amend the distribution of peak AKI severity and (4) the degree to which the hypothesised associative effects between AKI and final health outcomes, such as length of hospital stay, admission to ICU, need for RRT, 90-day mortality and risk of CKD, can be modified simply by amending the AKI distribution.

Model structure: follow-up Markov model

One potential route to patient benefit is that avoiding AKI or reducing its severity may reduce the risk of later developing CKD. As CKD is defined as a minimum of 3 months of persistent reduced renal function,^105,106 progression from AKI to CKD is incorporated into the Markov phase of the model.

Figure 21 illustrates the long-term follow-up model structure.

FIGURE 21

Markov model structure. Reproduced with permission from Jacobsen et al. This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and (more...)

After 90 days, the surviving cohort from each of the decision tree pathways enters a lifetime Markov model. The model follows a similar structure to those of Hall et al.⁹⁷ and Parikh et al.,⁹⁵ with six mutually exclusive health states: outpatient follow-up, CKD (stages 1–4), ESRD not requiring dialysis, ESRD requiring dialysis, post transplant and death. Members of the cohort enter the model either in the outpatient follow-up state, where they experience an annual baseline risk of developing CKD, or directly in the CKD state, with the proportion starting in the CKD state determined by the underlying CKD prevalence and the severity of AKI from the acute (decision tree) phase of the model. The base-case model assumes that the outpatient cohort have an increased risk of CKD in the first cycle that is dependent on their AKI experience, but, thereafter, the transition between the outpatient follow-up and CKD states is independent of whether or not a patient had AKI in the hospitalisation period. A sensitivity analysis explores the impact of an increased CKD risk applied for the full lifetime time horizon, as per Hall et al.⁹⁷

The cohort is then modelled to transition through the disease pathway, starting with CKD stages 1–4 (defined as a single Markov state), to ESRD, with or without the requirement for dialysis, the need for transplant, the success or failure of that transplant, and, ultimately, progression to death, with state-specific mortality probabilities. Those members of the cohort who experience transplant failure are assumed to return to the dialysis health state, in which the probability of a second transplant is the same as the probability of a first transplant. The cohort is exposed to a probability of all-cause mortality from each model state and assigned mortality probabilities based on the higher value of age- and sex-adjusted all-cause mortality or the disease state-specific mortality obtained from the literature.

The impact of an early Kidney Disease: Improving Global Outcomes care bundle on acute kidney injury

National-level guidelines¹⁶ indicate that, in a patient defined as being at risk of developing AKI through a positive biomarker result, all appropriate efforts should be made to ensure that AKI does not develop, and, if it does, it should be minimised in terms of severity (i.e. providing the maximum support possible for the kidneys). The model therefore assumes that all AKI patients will receive a KDIGO care bundle; the only difference between the testing strategies is the duration for which that bundle is implemented, with earlier implementation assumed to incur additional resource use in terms of fluid management, nephrologist review and pharmacist review of medications, as well as the removal of any potentially nephrotoxic agents when necessary.⁵

There are two potential mechanisms by which early adoption of a KDIGO care bundle might lead to patient benefit. These are to (1) avert AKI in people in whom it would otherwise develop and (2) shift the distribution of AKI severity (between KDIGO AKI stages 1–3), if AKI occurs.

Hall et al.⁹⁷ conducted a review of the literature to identify studies testing the impact of early preventative intervention for AKI. Their searches identified eight studies relevant to early intervention in the UK setting (excluding early RRT, which was deemed contentious). Four studies explored the impact of early nephrologist involvement, which was deemed to be the most reflective proxy for the non-specific care bundles that a patient may access as part of the KDIGO care bundle recommendations.⁵ The largest of these four studies, with a sample of 1096 participants, was used in the Hall et al.⁹⁷ economic model, and found that early nephrologist consultation reduced AKI incidence: adjusted odds ratio (early involvement vs. not) 0.71 (95% CI 0.53 to 0.95).

The EAG has conducted a supplementary targeted search of trials for the post-Hall et al.⁹⁷ period to identify any further potentially relevant studies exploring the impact of early preventative intervention or application of AKI care bundles on the probability of developing AKI and/or the severity of peak AKI. In brief, 39 additional titles and abstracts were identified from the targeted searches, of which 17 (44%) were full-text assessed. Based on the NICE scope,¹⁰⁰ KDIGO care guidelines⁵ and clinical expert opinion (Simon Sawhney, University of Aberdeen, 2019, personal communication), it was decided that studies testing the impact of a KDIGO care bundle provided the most appropriate source of data to populate the economic model. Three trials^110,116,117 (18%) assessed the effect of NephroCheck-guided application of a KDIGO bundle compared with standard care where information about the NephroCheck test result was not available to a patient’s hospital care team. No studies assessed the impact of NGAL-guided treatment.

All three studies reported results in terms of the probability of developing AKI.^110,116,117 However, only one study¹¹⁰ described the impact on both the incidence and severity of AKI. Meersch et al.¹¹⁰ reported the results of a single-centre trial, with a sample of 276 participants, in a German setting. All of the population had positive NephroCheck test results, using a 0.3 mg/dl threshold, consistent with the sources of diagnostic accuracy data obtained from the systematic review of diagnostic accuracy studies (see Chapter 3). Patients were then randomised to receive a strict implementation of the KDIGO guidelines or standard care. The intervention group included avoidance of nephrotoxic agents, discontinuation of angiotensin-converting enzyme inhibitors (ACEIs) and angiotensin receptor blockers (ARBs), close monitoring of urine output, close monitoring of serum creatinine levels, avoidance of hyperglycaemia (for 72 hours), consideration of alternatives to radiocontrast agents, and fluid optimisation. In the control (standard-care) group, Meersch et al.¹¹⁰ state that the recommendations of the American College of Cardiology Foundation 2011 were followed, including keeping mean arterial pressure at > 65 mmHg and central venous pressure at between 8 and 10 mmHg. ACEIs and ARBs were used only when the haemodynamic situation stabilised and hypertension occurred. It is unclear whether or not knowledge of the NephroCheck test result was revealed to the treating hospital team for patients in the standard-care arm of the study. The primary outcome in Meersch et al.¹¹⁰ was 72-hour AKI and the trial found an absolute risk reduction of 16.6% (95% CI 5.5% to 27.99%). The Meersch et al.¹¹⁰ study was supported by the German Research Foundation (Bonn, Germany), the European Society of Intensive Care Medicine (Brussels, Belgium), the Innovative Medizinische Forschung (Münster, Germany) and an unrestricted research grant from Astute Medical, Inc.

A second, smaller study (121 participants),¹¹⁶ also in a German setting, showed that NephroCheck-guided care demonstrated a trend towards a lower probability of AKI, although the results were not statistically significant, with an OR for standard care versus NephroCheck of 1.96 (95% CI 0.93 to 4.10). The study did, however, find significantly greater odds of AKI (defined as stage 2 and 3 combined) in the standard care group than in the NephroCheck group: OR for standard care versus NephroCheck 3.43 (95% CI 1.04 to 11.32). A third study,¹¹⁷ with only 100 participants, compared the effect of a NephroCheck-triggered consultation implementing KDIGO recommendations for AKI with the effect of standard care in an ED in Germany. AKI outcomes were similar in both groups. The probability of AKI stage 2 or 3 at day 1 post admission was 32.1% in the intervention group and 33.3% in the control group; at day 3, it was 38.9% in the intervention group and 39.1% in the control group. Neither the Göcze et al.¹¹⁶ study nor the Schanz et al.¹¹⁷ study reports any funding involvement from the test manufacturers.

As the Meersch et al.¹¹⁰ study has a larger sample, and reports data for both the probability of AKI and the distribution of AKI severity, given that it occurs, these data were used for the model base-case analysis. Although the clinical context of the immediate postoperative period after cardiac surgery from Meersch et al.¹¹⁰ is likely to be generalisable between the UK and other countries, the nature of the AKI insult (ischaemia/reperfusion, postoperative haemodynamic, oxidative stress, haemolysis, in people with cardiac comorbidity) is specific to this context, as is acknowledged by the authors. Accordingly, this study¹¹⁰ may not be generalisable to AKI in the context of other acute or critical illness circumstances in which biomarker performance and the potential for AKI prevention/mitigation may be different.

The model describes the potential impacts of a biomarker-guided care bundle on (1) the chance that patients may develop AKI and (2) the severity of AKI given that it occurs. The assumption is that early biomarker-guided implementation of the KDIGO care bundle may reduce the proportion of people who develop AKI and help ensure that, if they do develop AKI, it will be of reduced severity. These effects are applied probabilistically as RRs in the model for those with true-positive test results only, using log-normal distributions.

In the absence of any data on the impact of NGAL-guided KDIGO care bundles on the probability of developing AKI or the severity of AKI, the base-case model assumes that the potential to avert AKI is similar for both biomarkers. However, based on clinical expert opinion (Simon Sawhney, personal communication) and the manufacturer-described role of the tests, NGAL measures injury and can be used to define AKI, whereas NephroCheck can identify stress, thereby enabling intervention before AKI develops. Therefore, a sensitivity analysis explores a scenario in which the RR of AKI for NGAL-guided care is equal to 1, while retaining the same effect on the AKI distribution, given that AKI occurs as for NephroCheck. It is acknowledged that these assumptions are uncertain, and the sensitivity analysis may present a bias against NGAL if data were to become available to suggest an effect on AKI prevention.

It was assumed that there are no negative health effects of early intervention for the proportion of each test group with false-positive results, but the additional costs of the bundle were still incurred. The model also includes the functionality to explore the impact of an additional mortality risk, for example because of excessive resuscitation as a result of fluid administration or removal of effective, but nephrotoxic, treatments in patients with a false-positive test result.

Although the model describes the impact of biomarker-guided early intervention on the distribution of AKI, it is unclear whether or not these effects translate into final clinical and patient-relevant health outcomes, such as need for ICU care, need for RRT, mortality or the development of CKD. The limited evidence that exists from Meersch et al.¹¹⁰ suggests that, although there is a significant reduction in the primary study outcome of AKI within 72 hours for NephroCheck-guided implementation of a care bundle, compared with standard care (OR 0.483, 95% CI 0.293 to 0.796), this ability to avert AKI was not demonstrated to translate into improvements in a range of clinical and patient-relevant outcomes, including need for RRT therapy in hospital (OR 1.618, 95% CI 0.676 to 3.874), 90-day all-cause mortality (OR 1.213, 95% CI 0.486 to 3.028), ICU LOS (median difference 0 days, 95% CI –1 to 0 days) or hospital LOS (median difference 0 days, 95% CI –1 to 1 days). Although the study was not powered to detect differences in these outcomes, there are no trends in the data that are suggestive of an effect size. Furthermore, the uncertainty regarding the link between increased resource use and clinical outcomes is emphasised by Wilson et al.,¹¹⁸ who demonstrated in their RCT of an electronic alert system for AKI that an early warning system increases resource use (e.g. renal consultation), but with no evidence that this translates into measurable clinical or patient benefit in terms of mortality or LOS. Indeed, for a subgroup on a surgical ward, the mortality rate was significantly higher in the electronic alert group. As these causal links between AKI and changes in health outcomes are highly uncertain and hypothesised based on observational data in the model, extensive sensitivity analyses are conducted to test the impact of a range of plausible assumptions on cost-effectiveness.

Starting proportions applied in Markov cohorts

One plausible route to patient benefit from averting or reducing the severity of AKI is through the prevention of new CKD and the indirectly associated longer-term progression to ESRD and transplant. It should be noted that the model does not assume a direct effect of peak AKI on ESRD at 90 days; therefore, patients can enter the Markov model in either the outpatient follow-up or CKD (stages 1–4) health states only. A reanalysis of 2012 data from the Grampian cohort¹⁰² indicated that only a very small proportion [13/4314 (≈ 0.03%)] of patients with AKI, almost all of whom had underlying CKD, progressed directly to ESRD at 90 days. Therefore, we assumed no direct transition from the decision tree to the ESRD state in the Markov model. The starting proportions (after 90 days) for each health state are dependent on the decision tree pathway through which the cohort has come, and the peak AKI severity the cohort experienced. The baseline prevalence of CKD in the UK general population has been estimated from Kerr¹¹⁹ at 6.1%. However, in a group of critically ill, hospitalised patients, this prevalence may be substantially higher. For the base-case analysis, we use the underlying prevalence of CKD in the Grampian data set,¹⁰² calculated as the prevalence of CKD in all hospitalised patients having their kidney function monitored. Multiplying through by the sampling fraction for no CKD (20%) and taking the proportion of CKD/full sample gives the baseline prevalence in this group, calculated as (5935/53,691) × 100% = 11.05%.

Health-state transition probabilities

The baseline incidence of new-onset CKD for the Markov model uses the same source as Hall et al.,⁹⁷ with an annual probability of progressing from the outpatient state to the CKD state of 0.0044 (95% CI 0.0039 to 0.0049) for patients in the no-AKI cohort.¹¹¹ The data are obtained from a large cohort study of 97,782 ICU patients enrolled on the Swedish intensive care register. The parameter value 0.0044 reflects the CKD incidence at 1 year post ICU admission for the proportion of patients with no AKI. The same baseline proportion of CKD was applied for those without AKI and for those modelled to have had AKI averted as a result of early preventative treatment. The proportion of the no-AKI cohort starting in the CKD state at day 90 was calculated as the underlying prevalence plus the new annual incidence, adjusted to the 90-day time horizon of the decision tree component of the model.

Hazard ratios for AKI 1, AKI 2 and AKI 3 on the development of CKD (defined as CKD stage ≥ 3) were obtained from a systematic review by See et al.¹¹² The review included a total of 82 studies quantifying the association between AKI and longer-term renal outcomes (including CKD) and mortality. However, only three studies reported the impact of each stage of AKI on CKD development.^120–122 One study (104,764 participants) in a US setting generated slightly counterintuitive results, with point estimates of the HR reducing as AKI stage increased.¹²⁰ However, two other studies in Asian settings (with 77¹²² and 1363¹²¹ participants) illustrated an increasing HR for more severe AKI stages. The systematic review has meta-analysed these three studies; the summary effects by AKI stage on CKD, defined as CKD stage 3, are used in the base-case analysis. The advantage of these studies is that they allow a demonstration of the impact of adapting the distribution of AKI severity on longer-term development of CKD. However, they are not conducted in a UK setting and may lack relevance. An alternative source, reporting the HR for the association between AKI and CKD that is constant across all AKI stages, is reported by Sawhney et al.¹⁰² for 9004 hospitalised patients with AKI in Grampian. The HR for development of stage 4 CKD (AKI vs. no AKI) was 2.55 (95% CI 1.41 to 4.64). This study has the advantage of relevance to the setting, but does not include risks by AKI severity. However, it should be noted that the definition of CKD is stage 4 in Sawhney et al.,¹⁰² compared with stage 3 in the meta-analysed studies, which may limit the comparability of the reported HRs.

The HRs of CKD by AKI stage are applied to the new incidence over the first 90 days and to the first annual transition in the model. Thereafter, the transition probabilities from outpatient follow-up to CKD follow the baseline 0.0044 per year. This approach is based on expert opinion (Simon Sawhney, personal communication) that any longer-term effect of AKI on CKD development will become attenuated over time, particularly if it has not occurred in the first year following hospital discharge. A sensitivity analysis explores a scenario in which the HR of CKD is applied for the full duration of the model, reflecting the assumption applied in Hall et al.⁹⁷

Prevalence of CKD and incidence of new-onset CKD are parameterised in the model using beta distributions, and the HRs for the effect of peak AKI severity on CKD incidence (i.e. transition probabilities to CKD state) are parameterised using log-normal distributions.

Progression from chronic kidney disease

The transition probability from outpatient follow-up to CKD is 0.0044, as described previously. The model cohort can then progress from CKD to ESRD, with or without dialysis, and from ESRD to transplant according to the modelled transition probabilities. It is assumed that AKI can influence only the number of people who get CKD and then has no further direct effect on how fast they progress through the CKD stages to ESRD, dialysis or transplant. The cohort is also exposed to an increasing mortality risk as it progresses through more severe disease states from CKD (stages 1–4) to ESRD without dialysis and ESRD with dialysis. Transitions from CKD (stages 1–4) to ESRD, from ESRD (no dialysis) to ESRD (with dialysis), and from CKD (stages 1–4)/ESRD to death are obtained from Kent et al.,¹¹³ who reported data on progression of kidney disease from the large (7246 participants), international (Europe, North America and Australasia) Study of Heart and Renal Protection (SHARP) RCT. The median study follow-up was 4.9 years, participants had a mean age of 63 years and 64% of participants were male.

For those with ESRD on dialysis, the proportions transitioning to kidney transplant and mortality were obtained from 5-year data published in the 2018 UK Renal Registry report (table 1.17),¹¹⁵ which provided information on transition from incident RRT in 2012 to transplant and mortality 1, 3 and 5 years later. The 3- and 5-year probabilities were annualised; the 3-year probability was applied to years 2 and 3, and the 5-year probability was applied to year 4 onwards. These probabilities were converted to the relevant annual cycle-specific probabilities and applied in the model using tunnel states to track time from entering a given health state. The UK Renal Registry also provided data on the probability of transition back to dialysis for failed transplants and the probability of death over 5 years following transplant. After 5 years post transplant, mortality is assumed to revert to the general population all-cause mortality probability and the annual probability of transplant failure remains at that reported from years 3–5 in the UK Renal Registry. It is further assumed that the proportion of the cohort with a transplant failure return to dialysis, and that their probability of progressing from ESRD on dialysis to a second transplant is the same as the probability of progression to the first transplant.

In the first 5 years of the follow-up phase of the model, mortality in all Markov states is modelled as the average mortality risk for patients discharged from hospital and ICU, unless health state-specific (ESRD, dialysis or transplant) mortality is higher, in which case the latter is applied. If, at any point, mortality falls below all-cause mortality, all-cause mortality is applied in the model. The 5-year post-discharge mortality data were based on Lone et al.,¹¹⁴ a matched UK cohort study (mean age of 60 years) using national registries: the Scottish Intensive Care Society Audit Group, the Scottish Morbidity Record (SMR) of acute hospital admissions (SMR01) and the Scottish death records. The model base case used an average of the ICU and non-ICU cohorts. Beyond 5 years, patients in the outpatient follow-up health state had the age- and sex-adjusted all-cause mortality probability applied,¹²³ and those with CKD, ESRD, chronic dialysis or a transplant would be assigned the health state-specific mortality probability, unless the age- and sex-adjusted all-cause mortality was higher than the health state-specific mortality. A sensitivity analysis explores the impact of assigning long-term mortality risks that are dependent on whether or not the cohort had been admitted to ICU during the index hospitalisation.

Transition probabilities are incorporated into the model probabilistically using beta distributions. As the cycle lengths for the model in Hall et al.⁹⁷ are the same as the current assessment (annual), it was not necessary to provide any further adjustment of the published transition probabilities.

Diagnostic test costs

NephroCheck testing is usually conducted on an Astute140 Meter, costing £3000, and an additional meter would need to be purchased. This cost was converted to an annuity, assuming that the platform’s lifetime is 5 years and an annual depreciation rate of 3.5%. The test could also be conducted on a VITROS Immunodiagnostic System, although, currently, there is a limited installed base of these in UK hospitals,⁹⁷ which was confirmed at a NICE scoping workshop. The NGAL tests would not require a new platform for NGAL only, because it would be performed on platforms already available at the hospital laboratories. The capital costs of the laboratory analyser apportioned to each NGAL test are assumed to be negligible. The sensitivity analysis excludes capital and training costs to explore the impact on cost-effectiveness of scenarios in which a hospital might already have the required analyser in place and all staff are fully trained in its use.

The process of taking the sample for analysis, sending samples to the laboratory, processing at the laboratory and interpretation of test results would require the involvement of several members of the hospital team. A urine sample is first collected by a nurse, and then picked up by a porter, who takes it to the laboratory. It is assumed that, because the tests are classified as urgent samples, the porter would generally prioritise single test collection for the laboratory. A biomedical scientist conducts the diagnostic test in the laboratory. After completion of the test, the results from the laboratory would be authorised by a biochemist and released for review on the hospital information management system, where they can be interpreted by a nephrologist, an intensive care specialist or a junior doctor. The base-case analysis assumes an average of the three health-care professional costs for interpretation. Under some criteria (such as very abnormal results), a laboratory team might directly contact the care provider, but we assume that this approach would not be used routinely. For the purposes of test cost calculation, it is assumed that, on average, the role of interpreting the tests is equally split across the three specialist team members. The unit costs per hour for each of the staff resources involved in the testing process were obtained from PSSRU 2018:¹²⁴ to conduct the test – band 5 nurse (£37.00), porter costed as health-care assistant (£27.26) and band 6 biomedical scientist (£44.00); to interpret the test – medical consultant/nephrologist (£108) and junior doctor foundation year 2 (£32.00). It is assumed that a hospital consultant, nephrologist and junior doctor are all equally likely to be the health-care professional interpreting the test.

The duration of resource use for each member of the team is based on a combination of information provided by the manufacturer and clinical expert opinion (Simon Sawhney and Callum Kaye, personal communication) regarding the flow from obtaining the test sample to result interpretation. The staff time to process the test in the laboratory was based on the NICE request for information documents to the different test manufacturers and the final scope (NICE technical team, 2019, personal communication). Estimates of the time taken to prepare the urine sample and interpret the test were based on the EAG’s clinical expert opinion (Simon Sawhney and Callum Kaye, personal communication).

Four test strategies were compared in the economic model: NephroCheck, BioPorto urine NGAL, ARCHITECT urine NGAL and BioPorto plasma NGAL. The NGAL test manufacturer BioPorto has not identified costs separately by sample type (plasma or urine). It is therefore assumed that these tests incur equal costs. The cost of the Alinity i urine NGAL test was not considered in the base-case economic evaluation because the review identified no diagnostic accuracy data for the test. Full costing of each test is provided in Appendix 13, Table 31. Further details of maintenance costs and consumables for each test can be found in Appendix 13, Table 34.

Cost of early treatment

The additional cost of early treatment with the KDIGO care bundle was calculated as £106.36 per patient treated, assuming an additional 3 days’ application of the care bundle in test-positive patients. An additional 3 days of treatment was assumed in line with the primary outcome from Meersch et al.¹¹⁰ (i.e. AKI at 72 hours) and based on clinical expert opinion (Simon Sawhney and Callum Kaye, personal communication) that a care bundle could be implemented for up to an extra 3 days. The care bundle cost is based on the NICE guidelines for preventing AKI,¹⁶ which state that measures to prevent AKI are avoidance of nephrotoxic agents, discontinuation of medication (ACEIs and ARBs), close monitoring of serum creatinine and urine output, avoidance of hyperglycaemia, alternatives to radiocontrast and close haemodynamic monitoring. The NICE recommendations for preventing AKI¹⁶ include seeking advice from a nephrology team with regard to giving ‘iodinated contrast agent to adults with contraindications to intravenous fluids’ (© NICE 2013 Acute Kidney Injury: Prevention, Detection and Management. Clinical Guideline [CG169].¹⁶ Available from www.nice.org.uk/Guidance/CG169. All rights reserved. Subject to Notice of rights. NICE guidance is prepared for the National Health Service in England. All NICE guidance is subject to regular review and may be updated or withdrawn. NICE accepts no responsibility for the use of its content in this product/publication) and from a pharmacist with regards to medications (ACEIs, ARBs). Therefore, both nephrologist and pharmacist time are included in the cost of the care bundle. Further details of the cost calculation approach can be found in Appendix 13, Table 32. The additional cost of early adoption of the care bundle was applied to the proportion of the cohort with a positive biomarker test result, reflecting an assumption that care would be delivered for an additional 3 days over and above the cohort monitored using serum creatinine alone. The cost was applied using a gamma distribution with a standard deviation of 10% of the mean.

Acute phase costs

The base-case total cost in the acute phase (90 days) of the model depends on the number of days spent in hospital and the ICU. It also depends on the duration of acute RRT delivered to the proportion of AKI stage 3 patients receiving RRT. For the base-case analysis, data from the Adding Insult to Injury report show that 52% of RRT patients receive continuous RRT (daily) and 48% receive intermittent dialysis (an average of three sessions per week).³ The duration of RRT delivery is obtained from a randomised trial conducted in a US critical care setting¹²⁹ comparing intensive (6 days per week; n = 563) with less intensive (3 days per week; n = 561) RRT strategies. The mean duration of RRT per patient was similar in both groups: intensive, 13.4 days (SD 9.6 days); n = 563; and less intensive, 12.8 days (SD 9.3 days); n = 561. The base-case model conservatively assumes the less intensive duration for the application of costs in the economic model. To incorporate the uncertainty and to reflect the likely skewed nature of the distribution, the duration of RRT is incorporated probabilistically into the model using a log-normal distribution. Data available from an alternative source,¹³⁰ as used in NICE guidance for the comparison of early and late RRT, were not considered because median, rather than mean, durations were reported and the data were assessed as being of low quality in the NICE guidance.¹⁶ An additional daily excess cost of AKI was applied in a sensitivity analysis to capture the potential excess cost per day in hospital or an ICU of an AKI patient. This excess cost was not applied in the base-case scenarios because it was assumed that the cost of having AKI is captured in the cost of being in hospital or an ICU. All other acute costs and follow-up costs have a gamma distribution applied.

Long-term follow-up costs

There are four ways in which long-term follow-up costs may be driven by the proportion of the cohort that progress through different pathways from the initial decision tree. These are (1) whether or not long-term follow-up costs depend on whether or not a patient received ICU care in the initial decision tree, (2) whether or not there are additional follow-up costs beyond 5 years’ post index hospitalisation discharge, (3) whether or not an excess long-term cost is applied for the proportion of the cohort coming through AKI arms of the decision tree and (4) health state-specific costs incurred as the cohort progress through CKD stages to dialysis or transplant.

The outpatient follow-up costs in the Markov model post index hospitalisation discharge were obtained from Lone et al.,¹¹⁴ who reported 5 years of follow-up costs post index ICU and hospital discharge, using a matched cohort obtained from registries in Scotland [Scottish Intensive Care Society Audit Group, the SMR of acute hospital admissions (SMR01) and Scottish mortality data]. The base-case analysis assumes that the average of post-ICU and post-non-ICU admissions is applied in the Markov model. This is because patients in the cohort for this assessment are already deemed to be critically ill and at risk of needing ICU care, so might all be expected to have significant resource use post discharge. A sensitivity analysis allows the application of differential long-term costs that depend on whether or not a patient had received ICU care in the first 90 days.

The annual costs beyond 5 years are unknown. Therefore, the base-case analysis assumes no additional costs beyond year 5. A sensitivity analysis explores the impact of these assumptions by applying further costs between years 6 and 11 that reduce annually following a logarithmic function, with year 11 costs applied for the remaining duration of the model.

The base-case analysis assumes that there are no long-term excess follow-up costs as a result of having had AKI in the initial 90 days post hospitalisation. However, a sensitivity analysis explores a scenario in which patients entering the Markov model having had AKI in hospital incur an additional 15% of the non-AKI cohort costs for the first 5 years. The additional AKI cost factor was based on a proxy using the RR reported in Lone et al.¹¹⁴ on the number of admissions that patients on RRT had over 5 years, compared with those who were not on RRT. These additional costs are applied in the model as a sensitivity analysis, with a mean ratio of 1.15, log standard error 0.074, sampling from a log-normal distribution.

Annual cycle-specific health-state costs were applied to the proportion of the cohort transitioning through the CKD, ESRD, ESRD on dialysis and transplant health states. Costs were obtained from Kent et al.,¹¹³ using data from the SHARP trial reporting outpatient, day case and inpatient admissions. The CKD (stages 1–4) health-state cost applied in the model was calculated as the weighted average of CKD stages 1–3 and CKD stage 4, as reported in Kent et al.¹¹³ Therefore, the average weighted cost applied was £445.98 per cycle. The cost of medications (immunosuppressants for transplant patients, erythropoiesis-stimulating agents for dialysis patients and blood pressure medications for dialysis patients) were not captured in the study; therefore, these costs were added to the costs observed in Kent et al.¹¹³ The added transplant costs (immunosuppressants) were based on the approach applied in Scotland et al.¹³¹ for calculating the annual cost of immunosuppressants, using 2018 prices. The added costs to dialysis patients arising from blood pressure medications and erythropoiesis-stimulating agents are also based on the approach applied in Scotland et al.,¹³¹ with 2018 prices.

Critical care subgroup

For the critical care subgroup, the same parameter values as the all-comers are used for the downstream model probabilities, costs and utilities. This subgroup may be useful for decision-making as it could be considered as an alternative, potentially more seriously ill, definition of the population in the NICE scope. Although the group is defined as ‘critical care’, the populations described in the source diagnostic accuracy studies are often more reflective of a seriously ill patient group that would not yet be in ICU in the UK setting. The diagnostic accuracy data used for this subgroup are described in Table 22.

TABLE 22

Diagnostic accuracy data used for the critical care subgroup analysis

The results of the critical care subgroup analysis are provided in Table 23.

TABLE 23

Results of the critical care subgroup analysis

Cardiac surgery subgroup

Diagnostic accuracy data were not available from the systematic review for all biomarker strategies for the cardiac surgery group, and were available from only single studies for some tests. When data were not available from the review, we used pooled estimates from Hall et al.,⁹⁷ but note that this analysis should be considered with caution as it includes test manufacturers outside the scope of the NICE assessment. The diagnostic accuracy data for the cardiac surgery subgroup are provided in Table 24 and are included probabilistically in the model when possible.

TABLE 24

Diagnostic accuracy data used for cardiac surgery subgroup

Again, these results should be interpreted cautiously because of the lack of/limitations with the diagnostic accuracy data, and the questionable relevance of the downstream parameters/model structure for a cohort of post-cardiac patients only.

The results of the post-cardiac surgery subgroup analysis are provided in Table 25.

TABLE 25

Results of the post cardiac surgery subgroup analysis