Study selection
and summarise the initial and update review findings respectively. For clarity, the findings of the original and updated review are shown separately.
Preferred Reporting Items for Systematic Reviews and Meta-Analyses diagram of studies included in the initial systematic review of model-based economic evaluations.
Preferred Reporting Items for Systematic Reviews and Meta-Analyses diagram of studies included via the updated review.
The initial search strategy of the databases conducted in February 2015 identified 2671 records, of which 908 were duplicate records. Based on title and abstract screening, 74 reports were selected for full-text review, of which 20 studies were included (reported across 21 reports). One further study was identified via citation checking, giving a total of 21 studies reported across 22 reports.
The updated search strategy conducted in February 2020 identified 1357 records, of which 342 were duplicate studies. Based on title and abstract screening, 12 reports were selected for full-text review, of which 7 studies (reported across 7 reports) were included. No studies were identified via citation checking in the update review. In total, across the initial and updated searches, 28 studies were included in the review, reported across 29 reports.
Study characteristics
A summary of the included studies is provided in Table 43. Nine studies originated from USA, four studies were from Canada, three from the Netherlands, two from China, and one each from Australia, Europe, France, Germany, Iran, Japan, Korea, UK, Switzerland and Thailand.
Summary table of studies included in the systematic review of model-based economic evaluations
The majority of studies (n = 24) considered proteinuria or albuminuria tests: 8 specified the evaluated test(s) as reagent strip (or ‘dipstick’) tests;195–202 11 referred to urine albumin excretion, urine albumin concentration or urinary albumin-to-creatinine ratio tests;203–214 and 5 evaluated both dipstick and standard albuminuria tests.215–219 Six studies meanwhile evaluated eGFR testing: two of which incorporated both dipstick and eGFR tests; and one of which incorporated both albuminuria and eGFR tests.200,202,211,220–222 Of those studies evaluating eGFR-based strategies, four specified the equation underpinning the eGFR calculation: three used the MDRD equation, while one used the CKD-EPI equation. A final study evaluated a novel biomarker for CKD – described as a capillary electrophoresis–mass spectrometry-based urinary peptide classifier (CKD273).223
All of the identified studies evaluated screening strategies in at-risk groups or general populations not currently diagnosed with CKD. No studies were identified which assessed test-based monitoring strategies for patients with known CKD (i.e. matching the role of monitoring being considered in this HTA). Half of the studies (n = 14) focused on screening strategies for patients with diabetes195,201,203,204,210,215–219 or patients with diabetes and/or hypertension197,212,214,223 (both known risk-factors for CKD). Note in the initial 2015 review, these screening studies (conducted in diabetic and hypertensive populations) were listed as monitoring studies. In the updated review, it was decided that for clarity these studies should be listed as screening studies. Although repeated screening within high-risk populations can be classified as a type of early-stage monitoring for disease onset, in the context of this report (where the focus is on monitoring within patients already known to have CKD), we believe it is more appropriate to identify such studies as ‘screening’ studies. This terminology also aligns with that used by the original study authors. The other half focused on screening strategies in the general population,196,198,200,202,205–209,213,220–222 with one study considering school children,199 and a further study focusing on screening within a rural indigenous population.211 Of those studies focusing on population-based screening strategies, several also included subgroups or scenario analyses restricting screening to patients with diabetes and/or hypertension.196,198,202,207,209,221 Testing was most often administered in primary care,195–197,199,203,205,207–210,218,219 or had no clearly-defined setting.200,202,204,213,214,216,217,220,221,223 Less commonly explored settings included community care and secondary care settings.
The time horizons adopted in these studies were based either on the final age of the patient population or on a specific period of time. The vast majority of studies adopted long-term time horizons to capture downstream cost and health outcomes: 5 studies adopted a lifetime horizon until all the patients had died;195,202,203,210,221 10 studies ran the model analysis until the patient population had reached 90 years old or more;198,200,207–209,213,214,218,219,223 2 studies ran the analysis until the patient population had reached age 75 years;196,201 3 studies adopted a time horizon of 45–60 years;211,216,217 and 5 studies adopted a 25- to 30-year time horizon.197,204,210,212,215 Only two studies implemented time horizons of < 20 years.205,220 A further study did not incorporate a time horizon – this study adopted a decision tree approach and considered the cost-per-case of CKD detected for urine dipsticks targeting school-aged children.199 One final study failed to report the time horizon.222
Study methods
How was test diagnostic accuracy considered in the analysis?
Eight studies did not incorporate any measure of diagnostic accuracy into the model, despite evaluating test-based strategies.195,200,203,204,211,214,221,222 Of these, most did not mention or discuss the topic of accuracy at all. Only one explicitly reported their assumptions: in their analysis of annual screening for albuminuria in patients with type 2 diabetes, Golan and colleagues explained that the evaluated test for micro-albuminuria was assumed to have perfect diagnostic sensitivity in the model (which would favour the testing arm compared to the ‘treat all’ comparator), and that the outcomes for any FP cases resulting from testing were assumed to be captured within clinical trial data used to inform subsequent CKD progression rates within the model.203
Table 44 summarises the methods employed across the 20 studies that did explicitly incorporate diagnostic accuracy into their models. Fourteen studies incorporated the possibility of both FP and FN test results (typically using diagnostic sensitivity and specificity metrics).196,197,201,202,207,209,210,212,213,215,220,223 Six studies meanwhile only considered the possibility of FP test results (mostly modelled using either PPV or diagnostic specificity metrics);199,205,216–219 of those, three studies explained that FN test results were not considered since the evaluated test was assumed to have perfect diagnostic sensitivity218,219 or NPV,199 without providing any justification for this assumption, while two studies did not mention or discuss the possibility of FN results.216,217 A final study indirectly captured FPs by calculating the number of subjects needed to be screened and tested to identify and treat 1000 subjects with confirmed albuminuria.205
Summary of approaches to incorporating test diagnostic accuracy for studies included in the systematic review of model-based economic evaluations
In the majority of studies, diagnostic accuracy values were informed by a single published study;197–199,201,207–209,213,218,219,223 in two cases, the cited study was a systematic review.218,219 Two studies cited multiple publications informing the diagnostic accuracy estimates,196,220 one of which reported taking averages of the published values to derive the model diagnostic accuracy parameters.220 No studies reported conducting a formal literature review or meta-analysis to inform the model diagnostic accuracy values. Two studies derived accuracy values from a primary diagnostic accuracy study,212,215 and a further study derived values from insurance claims data.202 One study utilised individual-level data from available data sets of repeated test values, to construct linear random-effects models of longitudinal log-ACR values (for patients with type 1 and type 2 diabetes) incorporating: (1) between-subject variation in baseline ACR (i.e. the intercept); (2) the average change in ACR over time; and (3) the difference between observed ACR and the underlying ‘true’ ACR for an individual, described as the within-person variability in ACR measurement arising from assay variability and short-term biological variability.210 The FP and FN rates in this case were then calculated based on classification errors resulting in differences between the true and measured ACR values over different time points, based on the regression simulation models. Two final studies failed to report any source for the reported accuracy values.216,217
In terms of modelling the impact of test inaccuracies, patients with FN results were assumed to remain in ‘untreated’ health states, and thereby could not benefit from treatment for CKD (typically corresponding to ACE inhibitor or A2RB therapy) until future screening rounds. Most often the consequence of missed treatment was modelled as higher risks of progression to later CKD stages (which in turn could be associated with higher mortality risks and lower health-related quality of life values);196–198,201,202,207–210,213,220,223 two studies also included higher risks of cardiovascular disease (CVD) events for patients with FN results.198,202 One of the earlier studies identified quantified the impact of FN results in terms of patient lost ‘quality of life days’, based on an expert elicitation exercise conducted with 30 physicians.215
In the case of FP results, the most common consequence modelled was the unnecessary cost of additional/confirmatory testing.196,199,205,207–210,213,215,220 In most cases, confirmatory testing was assumed to have perfect accuracy, thus removing all FP cases at this stage (although this was typically not clearly reported). In several studies, FP cases were stated to undergo unnecessary treatment.197,201,210,212,216,217,223 However, in general it was not clear how long patients were assumed to remain on unnecessary treatment. In two studies, the consequences for FP cases were not reported.198,202
Of those studies that evaluated repeated testing scenarios and explicitly accounted for diagnostic accuracy in the model, the majority assumed that the same diagnostic accuracy values would apply over time.196–198,201,202,207–209,213,215–220,223 While no study explicitly stated this assumption, it may be inferred from the fact that single values of diagnostic accuracy were reported, and that no considerations concerning accuracy over repeated tests were discussed. Two studies meanwhile did attempt to account for changes in test diagnostic accuracy that could result from sampling at different time points.210,212 The study conducted by Wang and colleagues included a primary diagnostic accuracy study, in which the impact of within-patient biological variability and assay variability was explored by evaluating different repeated-sampling strategies over a 3-month period – consisting of different combinations of five samples taken at the first day ante meridiem of each month (labelled DAY-1, MONTH-2, MONTH-3), the second day in the first month (DAY-2) and a random spot urine sample taken in the afternoon of the first day (RANDOM-1).212 Farmer and colleagues meanwhile constructed random-effects linear regression models to simulate the trajectory of log-ACR values over time for two separate populations (type 1 and type 2 diabetes), which accounted for errors in measured ACR values resulting from within-patient biological variation and assay imprecision.210 Using these simulation models, the authors were able to model changing rates of FP and FN values over time.
What approach was used to model disease progression?
The majority of studies used a cohort Markov model approach to model disease progression. Four studies applied a microsimulation approach to a Markov model structure.207–209,213 This means that the model followed patients individually as they passed through different states of the Markov model rather than monitoring the average probability of events for a representative cohort of patients. An additional study, which included two separate cost-effectiveness models (one for each of type 1 and type 2 diabetes), utilised an individual based model structure in which a series of risk equations were run within each model cycle (dependant on individual patient characteristics), to predict the timing of a series of cardiovascular and mortality events, in addition to renal failure.210 Only one study did not model disease progression and instead used a decision tree approach to establish the cost-effectiveness of identifying cases of CKD amongst school children.199
Amongst those studies that described the progression of CKD (prior to ESKD), this was implemented through the use of progressive albuminuria195,201,203,204,210,214–219,223 or GFR states.202,222 In addition, four studies modelled CKD progression based on both albuminuria and GFR states:207–209,213 all of these were based on the ‘CKD Health Policy Model’, originally published by Hoerger and colleagues in 2010.206,207 Of those studies that tracked progression via albuminuria levels, health states were typically divided into ‘normal albuminuria’, ‘microalbuminuria’ and ‘macroalbuminuria’, before progression to end-stage disease states. Of those studies that tracked progression according to GFR levels, the stages of CKD were most often split into five GFR levels. Most studies assumed a step-by-step process of disease progression (i.e. patients could only move up one health state in the ladder of progression per model cycle), and only a minority explicitly allowed for the possibility of the reversibility in CKD severity.201,205,211
For the later stages of disease, the majority of studies included a single health state for ESKD, followed by death. Several others delineated this stage of disease, most often including states for ESKD with and without renal replacement therapy; and before and after kidney transplantation. In addition, beyond the modelled CKD health states, several studies also captured cardiovascular complications associated with CKD (including events such as ‘heart attack’, ‘stroke’, ‘cardiovascular disease’ and ‘coronary artery disease’) either in the form of separate health states or additional risks applied to CKD health states.195,198,202,205,207–210,213,223 One study also captured additional outcomes for patients with diabetes (blindness and amputation).210
Were the impacts of any delays on the testing and treatment pathway considered?
Since the focus of the included studies was on screening, the most common type of delay considered was in correctly identifying disease positive patients. The most frequently evaluated screening strategy involved annual screening.196–198,201,203,204,207–210,213–220,223 Several studies also explored increasing the screening interval beyond 1 year up to a maximum of 10 years, thereby implying that a patient who becomes eligible may have to wait up to 10 years before diagnosis and treatment is offered (although several studies allowed for the possibility of clinical presentations and diagnoses, or opportunistic screening, between scheduled screening intervals).207–210,213 Interestingly, only one study considered the possibility of a screening interval of < 1 year.195 It is noted that among the studies included in this review, screening interventions tended to be cost-effective among patients with diabetes and/or hypertension but were typically not cost-effective within general populations unless restricted to higher-risk groups or adopting a longer screening interval.
No studies considered the possibility of a delay in receiving results following testing. While it is acknowledged that for the modern healthcare systems considered in this review, these delays will be minimal, delays in the communication of abnormal test results to patients can still occur. In addition, receiving immediate treatment following a confirmed positive test seems to have been an implicit assumption made in all of the studies where testing and treatment were considered.
Which testing costs were captured in the analyses and how?
The cost of testing related to screening activities captured in the included studies was typically limited to the unit cost of the screening test alone.195,198,199,201–205,208,212,214–219,223 Eight studies also included the cost of a physician/GP visit associated with the initial screening test(s) undertaken.196,197,207,209–211,213,221 One study included additional screening costs (including transportation of equipment and personnel, advertisement, human resources and dissemination) related to undertaking screening within a remote indigenous population.211 One study provided no specific details regarding screening test costs,222 while another assumed that no screening costs would be incurred due to the fact that the tests evaluated were assumed to already be reported routinely in standard care.220 One study applied the cost of a physician visit only, without explicitly including the cost of the screening test.213
Of those studies that provided information on the included screening costs, the most commonly cited sources related to national costing tariffs such as the Medicare and Medicaid fee schedule or reimbursement rates in the USA,196,203,207,208,213,220 and similar costing resources across other jurisdictions.198,212,218,221 Two studies reported that the test costs were based on a ‘recommended retail price’223 or ‘notified fee’202 but did not provide specific sources for these costs. Four studies meanwhile reported that they based screening costs on local institution (e.g. hospital)199,204,214 or national215 cost data, but failed to provide specific details. Of the remaining studies, five based the costs on previous studies including cost-effectiveness models,205,209–211,219 while one derived costs from a survey of health service providers, based on the respondents quoted prices to add the evaluated screening test to standard care (specific details about what this cost was assumed to include were not reported).200 Five studies failed to provide a source for the included test costs,197,201,216,217,222 while one study stated that the reported test costs were based on assumed values.195
Six studies reported adopting a societal perspective,196,200–203,212 with the remainder using a healthcare provider/insurer perspective. Of those studies stating that a societal perspective was adopted, half of them did not provide any specific details.200,203,212 One study included societal costs relating to lost wages alone: Boulware et al. incorporated lost wages resulting from patients aged < 65 years unable to work in the ESKD health state.196 The two remaining studies captured additional elements of societal costs. Srisubat et al. included nonmedical costs relating to food and travel expenditure in addition to lost earnings, based on a cross-sectional survey of patients with normoalbuminuria, microalbuminuria, macroalbuminuria and ESKD;201 Go and colleagues based their model health state costs on a previously published costing study, which captured productivity loss, transportation and caregiver costs associated with hospital inpatient and outpatient visits.202 None of the identified studies specifically discussed societal costs resulting from having to attend testing (and re-testing) – rather the focus tended to be on societal costs associated with long-term treatment of disease.
