A cost-utility analysis was conducted for a UK (England) population in line with the NICE reference case [18]. This included adopting a National Health Service (NHS) and personal social services perspective for costs, measuring health benefits with quality-adjusted life years (QALYs), using a lifetime time horizon, and applying a 3.5% annual discount rate for costs and health effects. The incremental cost-effectiveness ratio (ICER) was estimated and cost effectiveness evaluated using the lower bound of the £20 000-30 000/QALY willingness-to-pay threshold range adopted by NICE [19]. Reporting was aligned with the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) checklist (Table S1) [20].
The analysis considered the cost effectiveness of ESG alongside LM versus LM alone based on the MERIT RCT with the model population limited to patients with class II obesity (i.e., who could potentially be eligible for ESG per current international and UK guidelines). In this MERIT subgroup (n = 115), the mean age at baseline was 42 years (range 20-64), mean BMI was 37.5 (range 34.33-39.91), and 14.0% of participants were male; a total of 58.1% had hypertension, 33.0% had type 2 diabetes, and 20.0% had sleep apnoea [21].
The comparison of ESG with LM alone in the model is relevant as MERIT is the only RCT conducted to date that has evaluated the effectiveness and safety of ESG and represents the most robust source of clinical evidence on the procedure. In the model, LM reflected Tier 3 weight management services recommended in NICE clinical guidelines [6]. This is typically administered over a 2-year duration and comprises a specialist physician, a dietician, a specialist nurse, and a clinical psychologist with access to physical therapy [22]. The ESG procedure was assumed to be performed using the same device (i.e., Overstitch™) and in the same outpatient setting as the MERIT study.
Authors JK, VM, SA, and BH provided expert clinical advice to validate the model structure and inputs/assumptions. Given the technical modelling nature of the study, patient group input was not solicited during model development, though results were presented to Obesity UK.
Model
A de novo 6-state Markov model was developed in Microsoft® Excel which included 5 BMI-based health states and an absorbing death state (Figure 1). Patients entered the model in the class II obesity state and could transition to the other model states depending on changes to weight and risk of death over the model time horizon (100 years minus the mean age of model patients at baseline). A cycle length of 6 months was used for the first year in order to reflect the immediate weight loss observed with ESG and annual cycles were used thereafter. A half-cycle correction was applied to account for the fact that events and transitions could occur at any point during the cycle.
Clinical parameters
Model clinical inputs and parameter values are summarised in Table 1 and Figure 2.
Model baseline characteristics (age and sex) and BMI group data required for the calculation of health state transition probabilities at 6 months, 1 year, and 2 years were based on patient-level data for the subgroup of patients with class II obesity from the MERIT study [21]. The last observation carried forward approach was used to impute missing BMI data at each timepoint, providing the last observation was ≤ 10 weeks before the timepoint (patients with last observations > 10 weeks before the timepoint were excluded). In the first year where 6-month cycles were used, in the event of missing data at week 26 (6 months), values were taken in order of preference from observations at week 24, week 30, or week 16, which were all monitored visits and considered sufficient to capture weight loss associated with ESG. The MERIT study had a 2-year follow-up duration with available data for LM alone limited to the first year before patients randomised to LM were permitted to cross over and receive an ESG. Consequently, assumptions were required to extrapolate transition probabilities over the remainder of the model time horizon. Weight loss was assumed to plateau after 2 years (with BMI remaining constant thereafter) for 80% of model patients receiving ESG. To account for the potential of weight regain following ESG, the remaining 20% of patients receiving ESG were assumed to gradually return to baseline BMI by 5 years based on a recent systematic review and meta-analysis of studies assessing weight regain following bariatric surgery [23]. Weight regain was assumed to occur in all patients receiving LM after 1 year with BMI gradually returning to baseline BMI by 5 years, consistent with the approach taken in NICE’s appraisal of liraglutide for the management of obesity [24].
Adverse events (AEs) included in the model were based on the incidence of any severe AE that occurred in the MERIT safety population comprising all study participants [14]. This resulted in the inclusion of abdominal abscess, upper gastrointestinal bleed, and malnutrition AEs for ESG, and no AEs for LM.
The MERIT study included a relatively small sample size, and no deaths were observed during the 2-year follow-up period. Mortality (Figure S1) was therefore estimated by applying BMI-specific mortality risks from a large UK population-based cohort study identified in a pragmatic literature search [25] to age/sex-matched general population mortality rates for 2021 from the Office for National Statistics [26]. The MERIT study similarly did not provide sufficient data to inform model inputs on the prevalence of obesity-related comorbidities. The prevalence of comorbidities in each health state was therefore estimated using BMI-specific rates identified through a pragmatic literature search. Comorbidities included in the model were type 2 diabetes [27], hypertension [28], sleep apnoea [29], non-alcoholic fatty liver disease [30], and gastro-oesophageal reflux disease [31]. Given the limited sample size in the MERIT study, no model subgroup analyses were conducted.
Utilities
EQ-5D is NICE’s preferred measure of health-related quality of life for informing cost-utility analyses [18]. SF-36 data were collected in MERIT and can be directly mapped to EQ-5D using a mapping algorithm such as that from Rowen et al, 2009 [32]. We therefore conducted an analysis of patient-level SF-36 data for the class II obesity MERIT subgroup using this algorithm to inform health state utility values (Table S2). A limited number of patients transitioned into the overweight health state (n = 28) and no patients transitioned into the healthy weight health state during study follow-up. As such, SF-36 data from MERIT was considered inadequate for deriving utility estimates for these health states; model parameter values (Table 2) were instead informed by a large UK population-based cohort study from Stephenson et al, 2021 identified in a pragmatic literature search on the association between BMI and quality of life [33]. Further, when directly mapped from the MERIT SF-36 data, the resulting utility estimates were higher than those reported by Stephenson et al, (Table S2), likely due to the ceiling effect [34]. Therefore, in line with NICE technical guidance [35], a linear mixed-effects model was used to estimate the incremental disutility associated with increasing BMI in the obesity I-III health states (Table S3). These disutilities were applied to the overweight health state utility value taken from Stephenson et al, to derive health state utility values for the obesity I-III health states (Table 2 and Table S4).
Disutilities were also applied for ESG-related AEs, with values for abdominal abscess, upper gastrointestinal bleed, and malnutrition (Table 2) identified through a pragmatic literature search [36-38]. Given the one-off nature of the procedure, these disutilities were applied in cycle 1 only.
To avoid double counting, comorbidity-associated disutility was assumed to be already captured in the BMI-based health state utility values.
Costs
Costs included in the model reflect intervention costs for both ESG and LM, costs associated with the management of AEs, and costs associated with treatment of obesity-related comorbidities (Table 2). These were based on 2020/21 unit costs where possible; older costs were inflated to 2020/21 values using the NHS Cost Inflation Index [39].
ESG costs were based on the cost of the device and hospital costs associated with delivery of the procedure. Costs for LM were applied to both treatment groups and were based on Tier 3 weight management, including healthcare professional visits, with cost categories (GP consultation, nurse consultation, dietician consultation, specialist consultation, consultation, and blood count) and frequency of visits taken from NICE’s appraisal of liraglutide [24]. Costs for clinical psychologist visits were also incorporated based on feedback from the clinical expert authors that these are routinely offered in Tier 3 weight management services. The cost of each component was sourced from Personal Social Services Research Unit 2021 unit costs and NHS England 2020/21 reference costs as applicable [39, 40].
As the model does not capture subsequent obesity treatment costs (e.g., bariatric procedures for eligible patients in whom treatment does not result in adequate or durable weight loss), LM costs were assumed to be incurred in both treatment groups for the duration of the model horizon. Although it is expected that a proportion of patients will not be compliant with LM medical advice about lifestyle and dietary changes over the duration of the intervention, a 100% compliance rate was assumed in the absence of robust data.
Costs for the management of obesity-related comorbidities were based on annual costs identified through a pragmatic literature search [27, 29-31, 41-44]. These annual costs were combined with the previously described comorbidity prevalence rates to estimate the total comorbidity cost for each treatment per health state per model cycle (Table S10). Costs for the management of severe AEs were applied as one-off costs in cycle 1 and sourced from the National Cost Collection 2020/21 [40].
Sensitivity and scenario analyses
Deterministic one-way sensitivity analyses (OWSA) were conducted for each model parameter across ranges equal to the 95% confidence intervals. These were mostly calculated using a standard error of ± 20% of the mean value for each parameter. Exceptions were health state utilities for which the values were sourced from the literature (for the healthy weight and overweight health states) and calculated (for the remaining health states), and the mortality hazard ratio for which the standard error was calculated. Results were plotted on a Tornado diagram to identify key drivers of cost effectiveness.
A probabilistic sensitivity analysis was conducted using 10 000 iterations to characterise overall uncertainty in the cost-effectiveness results, with values for each parameter simultaneously drawn from their individual uncertainty distribution. Results were plotted on an incremental cost-effectiveness plane scatter plot to visualise uncertainty and a cost-effectiveness acceptability curve was generated to show the probability of ESG being cost effective over a range of willingness-to-pay thresholds (£0-50 000/QALY).
A full list of model parameters including uncertainties and distributions is provided in the supplementary materials.
Scenario analyses were conducted to explore structural uncertainty related to important model assumptions/inputs including use of alternative long-term BMI extrapolations. Additional scenario analyses included use of an alternative mapping algorithm from Ara and Brazier, 2008 [45] and with all health state utility estimates based on values reported by Stephenson et al.