Identification of studies

The following bibliographic databases were searched: MEDLINE (via Ovid), MEDLINE In-Process & Other Non-Indexed Citations (via Ovid), EMBASE (Ovid), Web of Science (via Clarivate Analytics), NHS Economic Evaluation Database (NHS EED) and HTA (the Cochrane Library) and EconLit (EBSCOhost). In addition, searches were carried out on the following websites: Health Utilities Database (HUD) [School of Health and Related Research (ScHARR)] (URL: www.scharrhud.org/; accessed 7 August 2018), Health Economics Research Centre (HERC) (Oxford) (URL: www.herc.ox.ac.uk/publications; accessed 7 August 2018), EQ-5D (EuroQol) (URL: https://euroqol.org/search-for-eq-5d-publications/; accessed 7 August 2018), Cost-effectiveness Analysis Registry (URL: https://cevr.tuftsmedicalcenter.org/databases/cea-registry; accessed 7 August 2018) and International Society for Pharmacoeconomics and Outcomes Research (ISPOR) (URL: www.ispor.org/; accessed 7 August 2018).

The searches were developed and run by an information specialist (SR) in July 2018 and updated in November 2018. They comprised terms for RA, terms for TNF-α inhibitors and terms for ELISA testing. Search filters were used to limit the searches to cost-effectiveness studies. No date or language limits were used.

Separate searches were also carried out for appropriate costs and health utilities using a variety of search terms and filters. These searches were carried out in several iterations to look for different aspects of costs and health utilities for RA and ELISA tests, as needed.

The full search strategies for each database, for cost-effectiveness, and one example iteration of the utility searches are provided in Appendix 1. The database search results were exported to and deduplicated using EndNote X7. Deduplication was also performed by manual checking. Screening was carried out independently by two reviewers. Disagreements between reviewers were resolved by consensus. All of the references that were considered for inclusion by either reviewer at the title and abstract stage were included for full-text screening.

Eligibility criteria

Studies eligible for inclusion in the systematic review were selected according to the inclusion and exclusion criteria outlined in a population, intervention, comparator, outcome (PICO) template. The inclusion criteria for population, interventions and comparator were as described in Chapter 2, Population, Interventions and Comparator. The following types of economic evaluation were included: cost–utility, cost-effectiveness, cost–benefit, cost–consequences and cost-minimisation analyses. Systematic reviews of economic studies were also considered.

Data extraction

Study characteristics and results were extracted and summarised by one reviewer (MRH). The evidence was assessed using narrative synthesis supported by summary data extraction tables.

Critical appraisal

The quality of the selected studies was evaluated by one reviewer in accordance with the Consensus on Health Economic Criteria (CHEC).⁴⁹ Studies based on decision models were further quality assessed using the checklist developed by Philips et al.⁵⁰

Non-model-based studies

The two studies, by Ucar et al.⁴² (INGEBIO) and Pascual-Salcedo et al.,⁴⁴ were reported as abstracts. Ucar et al.⁴² investigated the impact of monitoring levels of ADL and anti-ADL antibody in people with RA, PsA and AS on the annual direct costs to the health system and health outcomes compared with conventional practice in Spain. The economic analysis reported in Ucar et al.⁴² was based on clinical outcomes from a pragmatic, non-randomised, non-inferiority study. Trough levels of ADL and anti-ADL antibody were measured with Promonitor-ADL and Promonitor-ANTI-ADL. Physicians were not obliged to adhere to any therapeutic algorithm when making treatment decisions for participants in the intervention group. In the control group, treatment decisions were based on clinical judgement only. A total of 169 people were recruited, of whom 63 (37.3%) had RA (30 in the intervention group and 33 in the control group). Ucar et al.⁴² reported the result for all participants and did not report results by subgroup (disease categories).Therefore, it is difficult to generalise the results to people with RA. The mean cost of ADL (Humira^®) treatment per patient-year and the mean quality-adjusted life-years (QALYs) accrued over the observation period in the intervention and control arms were estimated. The authors reported, compared with the control group, those in the intervention group had better quality of life and lower risk of flares, and incurred lower treatment costs. The average ADL acquisition cost per patient-year was €10,664.54 and €9856.45 in the control and the intervention arms, respectively (–€808.08, 8% savings); the results were reported for the total (mixed) population. Given that the study was available in abstract form only, it was not clear how the mean QALYs were calculated.

The Pascual-Salcedo et al.⁴⁴ study aimed to compare the clinical and economic impact of TDM, based on trough serum drug levels, in people with RA and spondyloarthritis (SpA) in remission or with LDA. This was an observational study of routine clinical practice. The study included a total of 88 participants (RA, n = 43; SpA, n = 45) who were treated with three TNF-α inhibitors (IFX, n = 31; ADL; n = 29; and ETN, n = 28). Participants were followed for 7 years (2006–12). Drug levels were measured using ELISA tests. No further information on the test was given in the abstract. For each participant two time periods were examined: before and after the introduction of TNF-α drug monitoring (2006–9 and 2010–12, respectively). All participants in this study had stable clinical activity in both time periods. Pascual-Salcedo et al.⁴⁴ reported the monthly value of spared drug to be €91.62 per participant treated with IFX (assuming a mean participant weight of 70 kg), €324 per participant on ADL and €257 per participant on ETN.

Model-based studies

Three model-based economic evaluations were identified in the systematic review (see Appendix 6,Table 50). All were conducted in Europe (the Netherlands, Finland and the UK).

Krieckaert et al.⁵³

Krieckaert et al.⁵³ conducted a cost–utility study that investigated the measurement of ADL levels in people with RA. ADL levels were measured using in-house ELISA (Sanquin) for 3 years in a cohort of 272 ADL-treated participants with RA recruited at the Department of Rheumatology, Jan van Breemen Institute, Amsterdam.⁵⁵ These participants were compared with a cohort of 1034 participants from the Utrecht Rheumatoid Arthritis Cohort (URAC), who received other treatments based on clinical judgement. The clinical characteristics of these participants are not clearly discernible from the cited references. Participants in the intervention cohort were tested after 4, 16, 28, 40 and 52 weeks of treatment and every 6 months thereafter. However, in the economic analysis, the authors modelled ELISA testing at 28 weeks only. After 3 years, 76 of a total 272 participants (28%) developed anti-ADL antibodies; 51 (67%) participants developed these during the first 28 weeks of treatment. Over the course of the study, participants with measurable antibody levels were 3.6 times less likely [hazard ratio (HR) 3.6, 95% CI 1.8 to 7.2; p < 0.001] to revert to minimal disease activity (defined as a DAS28 of < 3.2) and 7.1 times less likely (HR 7.1, 95% CI 2.1 to 23.4; p < 0.001) to enter sustained remission (DAS28 of < 2.6). Clinical outcomes from Bartelds et al.⁵⁵ are summarised in Table 19.

TABLE 19

Clinical outcomes from Bartelds et al.

Of note, this study⁵⁵ was excluded from the clinical effectiveness systematic review because it did not meet the inclusion criteria for the population (the population was treatment naïve and had active disease). As this paper was excluded at the first-screening stage (titles and abstracts), it was not included in the list of excluded studies (see Appendix 2).

Krieckaert et al.⁵³ modelled a treatment algorithm (see figure 1 in Krieckaert et al.⁵³) based on TDM in RA patients. The authors used a Markov model with 3-month cycles and a time horizon of 3 years using microsimulation for analysis. The analysis was performed probabilistically. Discounting was applied at 4% for costs and 1.5% for utilities, following the Dutch national guidelines.⁵⁶ Results were reported from both health-care and societal perspectives. The Markov model health states were based on categorisation of DAS28 as below:

• remission (DAS28 of < 2.6)
• LDA (DAS28 of ≥ 2.6 and < 3.2)
• MDA (DAS28 of ≥ 3.2 and ≤ 5.1)
• high disease activity (HDA) (DAS28 of > 5.1).

Transition probabilities were estimated using a regression function that was derived from the URAC cohort outcome data.⁵⁵ Costs included direct medical and productivity costs. Utility was calculated based on the EQ-5D classification outcomes recorded in the URAC study.

Testing with ELISA kits was cost-saving from both the societal and the health-care perspective (Table 20). The test-based treatment strategy resulted in lower costs (due to the reduction in the treatment cost) and greater QALYs.

TABLE 20

Cost-effectiveness results reported in Krickaert et al.

A probabilistic sensitivity analysis around the base-case scenario predicted that ELISA testing would dominate usual care in 72% of scenarios. Scenario sensitivity analyses around, for example, the drug level cut-offs used, or the definitions of a good EULAR response, showed that ELISA testing is generally cost-saving, although some scenarios reported loss of QALYs.

Gavan¹⁷

Gavan¹⁷ evaluated the cost-effectiveness of using ELISA testing to monitor people with RA who are treated with ADL. In total, 12 different ELISA-based strategies were compared with the current practice in England (i.e. no TNF-α testing) (see Table 51 in Appendix 7). These strategies were a combination of using monitoring tests during response to the drug and after remission. The author considered a frequency of testing of every 3 or 6 months in responders to therapy. Among patients in remission, testing was considered after 2 and 3 years’ remission.

In Gavan,¹⁷ four lines of treatment were modelled, as shown in Figure 2.

FIGURE 2

Service pathway of RA treatment in England. Adapted from Gavan.

A discrete-event simulation modelling approach was used. The following competing events were considered: time to death, ADL failure, rituximab (Rituxan^®; Roche, Basel, Switzerland) failure, tocilizumab failure, time to development of antibodies against ADL, remission, EULAR response and HAQ progression. The model simulated 20,000 hypothetical patients, representative of the population with RA in England, using summary attributes of patients from the BSRBR-RA.

One of the test strategies considered in Gavan¹⁷ was monitoring drug and antibody levels in participants who were responding to treatment to avoid the harm associated with secondary non-response. Another possible test strategy was dose adjustment in patients in remission, informed by the results of TNF-α testing. Figure 3 shows the algorithm used by Gavan¹⁷ for management decisions in participants who had TDM performed.

FIGURE 3

Algorithm for test interpretation used in Gavan. Adapted from Gavan.

Utilities were calculated by mapping the HAQ score from the BSRBR-RA, by using a quadratic mapping algorithm estimated previously for the NICE TA195 by Malottki et al.⁵⁹ Costs included the costs of treatment, hospitalisations and testing. Quantities of resource utilisation were derived from published sources^34,37 and unit costs were taken from the NHS Reference Costs 2015–16⁶⁰ and the BNF.²¹

Based on the 12 strategies that were modelled (see Appendix 7, Table 51), Gavan¹⁷ concluded that routine use of ADL testing was cost-effective compared with current practice, but was unlikely to be cost-effective relative to dose reduction (without testing) for people in remission (strategy 11). Compared with current practice, strategies 1–6 and strategy 8 were estimated to be cost-effective. Strategies 9 and 10 were estimated to be less costly, but produced a lower QALY gain than current practice. Strategy 7 was dominated by current practice, that is current practice was associated with lower costs and a higher QALY gain than strategy 7. In the incremental analysis, all but three strategies (strategies 1, 3 and 11) were shown to be dominated or extendedly dominated by another strategy, that is that another strategy or combination of strategies was cheaper and produced more QALYs. Of the three remaining strategies, strategies 1 (testing levels of ADL and anti-ADL antibodies every 3 months) and 3 (testing levels of ADL and anti-ADL antibodies every 3 months, testing ADL level in patients in remission after 2 years) were not cost-effective compared with strategy 11 (no testing, halving ADL dose in all patients after responding for 2 years, irrespective of their drug level) at a willingness to pay of £20,000–30,000 per QALY gained: the incremental cost-effectiveness ratio (ICER) for strategy 1 versus 11 was £38,575 per QALY and the ICER for strategy 3 versus 11 was £37,043 per QALY. Given that strategy 11 consists of dose reduction after 2 years for people in remission, the analysis of the chosen strategies suggests that ADL testing may not be cost-effective compared with dose reduction alone.

Discussion

A systematic literature search performed in July 2018 and updated in November 2018 identified five publications that were relevant to the decision problem, with two of these available in abstract format only. Furthermore, only two (out of six) TNF-α testing kits from the NICE scope⁴⁸ (Promonitor and Sanquin) and three (out of five) TNF-α inhibitors (ADL, ETN and IFX) were considered in the selected studies (Table 21).

TABLE 21

Cost-effectiveness evidence relevant to specific combinations of TNF-α inhibitors and test kits from the NICE scope

Both Krieckaert et al.⁵³ and Laine et al.⁵⁴ concluded that TDM was cost-saving compared with standard care, based on follow-up periods of up to 3 years. Krieckaert et al.⁵³ reported a formal cost-per-QALY analysis in which TDM dominated standard care in the base-case scenario in 72% of simulations. The ICERs are arguably somewhat meaningless given the small incremental QALYs involved. In a range of sensitivity analyses, a net loss of QALYs with respect to the intervention was associated with drug-level cut-off points, the use of EULAR good response as an outcome or the use of biologicals other than TNF-α inhibitors. With regard to UK clinical practice, Krieckaert et al.⁵³ modelled testing at the 28th week from treatment initiation and considered dose reduction by prolongation of the interval between drug administrations in responders with high levels of ADL (> 12 mg/l). However, in the UK there are variations in when treatment decisions in people with RA who are treated with biologics are made. In responders to TNF-α inhibitors, decisions could either be made 9–12 months after treatment initiation or be adjusted approximately 2 years after the initiation of biological therapy. However, in non-responders, testing may be considered earlier to detect whether non-response to biologics is due to low drug levels or the presence of anti-drug antibodies.

Laine et al.⁵⁴ did not report a cost per QALY analysis, although they attempted to analyse the frequency and cost impact of non-testing with regard to inappropriate treatment decisions (e.g. continuation of ineffective therapy). It was assumed that participants in the routine practice arm would typically experience 3 months’ delay in receiving optimal treatment compared with participants in the intervention arm. This was justified based on the typical follow-up intervals of participants in Finland. Of note, in Finland anti-drug antibody levels of at least 30 AU/ml (arbitrary unit/ml) rather than 12 AU/ml are considered clinically significant.

Both the INGEBIO study⁴³ (n = 169) and Pascual-Salcedo et al.⁴⁴ (n = 88) recruited mixed populations in which, respectively, only 37% and 50% of participants had a diagnosis of RA. In addition, only limited details of the input parameters and analysis were provided, specifically:

• No details of utility values or incremental QALY outcomes were provided.
• The studies did not consider specific test-based treatment algorithms.
• Pascual-Salcedo et al.⁴⁴ did not specify which ELISA kits were used in their study.

Furthermore, the allocation of participants to groups in the INGEBIO study was site dependent and physicians were not obliged to follow any particular algorithm with regard to treatment. However, the statistical analysis plan was not documented and the assumption of independence of observations may not be appropriate. Therefore, the statistical significance of the reported results may be insecure.

The recent study by Gavan¹⁷ perhaps most closely matches the decision problem. In this study,¹⁷ modelling was based on patient data from the BSRBR-RA register, which is the main source of evidence on the use of biologics in people with RA in the UK. Furthermore, the research questions addressed in Gavan¹⁷ are most relevant to the decision problem considered in this report. However, Gavan¹⁷ did not consider any specific test kit and only ADL treatment was modelled as the first line.

Gavan¹⁷ considered three research questions, namely:

1. What is the existing economic evidence for stratified medicine in RA?
2. How are decisions on treatment with biological therapies for patients with RA in current practice made in England?
3. Are treatment decisions stratified by levels of ADL and anti-ADL antibody in patients with RA in England a relatively cost-effective use of health-care resources?

Research questions 1 and 2 have, to some extent, also been addressed by the searches and consultations for this review. However, Gavan¹⁷ considered any strategy involving biomarker testing to stratify decisions regarding treatment with any pharmacological therapy, whereas the current review focuses on the use of ELISA to monitor response to TNF-α inhibitor treatment. Research question 3 aligns closely with the decision problem for this appraisal. Although Gavan¹⁷ pointed out that there was a high degree of decision uncertainty and reported an expected value of perfect information estimate of £7M, the decision uncertainty was based around the cost of testing and test accuracy.

Based on these searches, it is clear that further exploration of this question, including de novo modelling, would be appropriate. Sufficient prior evidence with regard to the entirety of the decision problem and/or UK populations on which to base decision making clearly does not exist, especially given discrepancies in the conclusions of the studies presented. Of these studies, only Gavan¹⁷ could be considered to be of sufficient quality; however, no evidence was identified with regard to the use of test kits for either CTZ or GLM treatments, and no studies evaluating IDKmonitor ELISA kits, LISA-TRACKER ELISA kits, RIDASCREEN ELISA kits or MabTrack ELISA kits were identified.