Clinical trial design
The FIT trial was a randomised, controlled, assessor-blinded intervention trial, which investigated the effectiveness and costs of twelve monthly, 5-consecutive-days FMD cycles as an adjunct to usual care in patients with type 2 diabetes [14]. The trial was conducted at the Leiden University Medical Centre (LUMC) in the Netherlands, between 20 November 2018 and 5 August 2021. The trial was performed according to the principles of the Declaration of Helsinki, in accordance with the Medical Research Involving Human Subjects Act, and to the standards of Good Clinical Practice. The Medical Research Ethics Committee of the LUMC approved the protocol and amendments. All participants provided written informed consent before entry into the study. The study was registered at ClinicalTrials.gov: NCT03811587. Registration was initiated prior to the start of the trial, but due to a delay within the registration process, online publication occurred after the start of the trial.
Participants and intervention
Participants were recruited from general practice centres. They could be included if they were diagnosed with type 2 diabetes, had a BMI ≥ 27 kg/m2, were aged > 18 years and < 75 years, and treated with lifestyle advice alone, while their HbA1c was above 48 mmol/mol, or with lifestyle advice plus metformin as the only glucose-lowering drug regardless of their HbA1c. After randomisation, participants in the FMD group received twelve 5-consecutive-days meal-replacement FMD cycles every month for one year as an adjunct to usual care by their primary care providers. The control group received usual care only. The FMD consisted of complete meal replacement products, which included mainly soups, bars and snacks. The first day provided approximately 4600 kJ (1100 kcal; 10% protein, 56% fat and 34% complex carbohydrate); days 2–5 were identical and provided approximately 3150 kJ (750 kcal; 9% protein, 44% fat, 47% complex carbohydrate). Further details of the study design, intervention and exclusion criteria can be found in the study protocol [14]. One hundred patients with type 2 diabetes were enrolled in the FIT trial, and 92 participants completed baseline measurements [15]. Data from these 92 participants is used for the cost-effectiveness analysis. Further details on enrolment, allocation, reasons for drop-out, follow-up and effects on metabolism and anthropometrics are described elsewhere [15].
Cost-effectiveness analysis
A cost-effectiveness analysis was performed of the FIT trial data, comprising two parts. Firstly, we conducted a trial-based analysis evaluating the costs and benefits of the FMD programme as an adjunct to usual care as compared to usual care alone over the trial period of twelve months. Second, the results of the FIT trial were extrapolated to estimate the Lifetime cost-effectiveness of adding an FMD to usual care, using the United Kingdom Prospective Diabetes Study Outcomes Model version 2.2 (UKPDS-OM2.2) [18]. The main analysis was performed from a healthcare perspective. Analyses were conducted using STATA version 17 for Windows. Figures were created in GraphPad Prism version 9.0.1 for Windows.
Trial-based analysis
A cost-utility analysis combined health outcomes and costs of care provided during the twelve-month follow-up of the FIT trial to yield short term results, which were compared between the two study arms.
Quality-adjusted life-years
Data on participants’ quality of life was gathered at baseline, and after three, six, nine and twelve months follow-up, using the EQ-5D-5L questionnaire [19]. The Dutch EQ-5D-5L tariff was used to calculate utility values at each measurement [20]. These were used to calculate quality-adjusted life-years (QALYs) for the study duration by calculating the area under the utility curves for each individual patient.
Costs
Costs were calculated based on questionnaires filled out by participants at three, six, nine, and twelve months follow-up (Cost questionnaire in Supplementary Material). The self-designed questionnaire pertained to usage of healthcare services related to type 2 diabetes. Medication use was gathered at baseline, six and twelve months. The costs of the FMD programme per participant were calculated by multiplying the costs per day with the number of days that the FMD was actually consumed by that participant. Healthcare usage was valued using Dutch reference unit prices from the Dutch manual for costing research and consumer market prices [21]. All prices were converted to 2023 values using the Dutch national consumer price index [22]. Additional information on cost calculations can be found in the Supplementary Material (Additional Methods and Table S1).
Statistical analysis
All analyses were conducted following an intention-to-treat principle. For missing data, multiple imputations were performed with a total of 100 imputed datasets (Additional Methods in Supplementary Material). The 100 datasets were analysed separately, and estimates for utilities and costs were pooled using Rubin’s rules [23].
Bootstrapping with 1000 replications was used to estimate the statistical uncertainty surrounding the cost-effectiveness after adjusting the healthcare costs and QALYs for baseline values, using seemingly unrelated regression, which accounts for the correlation between costs and effects. Variables were taken into account as a confounder, if the estimated regression coefficient for the cost or effect differences changed by 10% or more [24]. Bootstrapped differences in cost and effect were plotted on cost-effectiveness planes. Subsequently, the cost-effectiveness of using the FMD (in addition to usual care) was compared to usual care only following a net benefit (NB) approach [21]. An intervention is considered cost-effective if the NB is higher for a given willingness to pay (WTP) threshold. The NB estimate per group was calculated by multiplying the adjusted difference in mean QALYs with WTP per QALY, and then subtracting the adjusted difference in mean cost between FMD compared to usual care. For each WTP value, the fraction of the bootstrap replications for which the NB of the FMD group was higher than the control group is shown in the cost-effectiveness acceptability curve. This curve reflects the probability of the intervention being cost-effective over a range of threshold values for willingness to pay for a QALY. A range of WTP values from 0 to 80,000 euros per QALY was used. At the probability of 0.5, there is no preference for either of the treatments.
The main analysis was performed from a healthcare perspective, using healthcare costs only and the QALYs calculated from the EQ-5D-5L. Two additional sensitivity analyses were performed: using QALYs calculated from the EQ-VAS instead of the EQ-5D-5L [1] and adopting a societal perspective instead of a healthcare perspective [2]. To this end, scores of the EuroQol Visual Analogue Scale (EQ-VAS), which was administered together with the EQ-5D-5L, were rescaled to a range of 0–1 and used to calculate QALYs using the area-under-the-curve method. For the societal perspective (regarding the second sensitivity analysis), costs consisted of healthcare costs, food costs and costs of productivity loss due to absence from work. Since the FMD is a meal replacement programme, total costs of food for the entire year were assessed. Food costs were determined with the presumed market price from the National Institute of Budget Education in the Netherlands, taking household sizes into account (Additional Methods in Supplementary Material) [25]. Costs of productivity loss were based on patient reports of absenteeism from paid and unpaid work at three, six, nine and twelve months follow-up. Costs of absenteeism of paid work were estimated with the friction cost method, using a friction period of 12 weeks, and hourly wages according to the Dutch manual for costing research [21, 26]. Costs of absence of unpaid work were assessed by multiplying the number of hours of unpaid labour lost by the average gross hourly wage of a domestic worker [21].
Lifetime model-based analysis
For the Lifetime cost-effectiveness analysis from a healthcare perspective, the United Kingdom Prospective Diabetes Study Outcomes Model version 2.2 (UKPDS-OM2.2) was used [18]. This model was used to extrapolate outcomes for individual FIT trial participants from end of trial (twelve months) until death (hereafter called post-trial results). The UKPDS-OM2.2 is based on data from the UKPDS trial and its ten-year post-trial monitoring [27]. It can be used to predict the occurrence of diabetes-related complications over a lifetime as well as life expectancy, and estimate healthcare costs and QALYs [18].
To simulate outcomes and estimate costs beyond the trial period, participant data influencing cardiovascular risk was entered from the end of study follow-up into the UKPDS-OM2.2. Variables included were age, duration of type 2 diabetes, glycated haemoglobin (HbA1c) and history of cardiovascular disease among others (Additional Methods in Supplementary Material). Missing data was imputed in the within-trial analysis, and the 100 imputed datasets were used as input in the UKPDS-OM2.2 model. In UKPDS-OM2.2, risk factors can be updated yearly, using the build-in equations that estimate risk factor progression over the years [28]. The mean utility based on the EQ-5D-5L from the end of follow-up of the FIT trial was entered into the model. Yearly treatment costs for type 2 diabetes and costs for complications were based on literature data (Table S2). Costs were discounted at 4.0% per annum, and QALYs at 1.5% per annum according to the Dutch guideline for economic evaluations [21] to account for the fact that people generally value future costs and effects less than current costs and effects. All prices were converted to 2023 values using the Dutch national consumer price index [22]. When running the UKPDS-OM2.2 model, uncertainty was taken into account by running 1000 Monte Carlo loops to reduce first order uncertainty and 100 sets of parameter estimates to address second order uncertainty.
Within-trial results on healthcare costs and QALYs were added to the post-trial results on an individual participant level, resulting in Lifetime estimations. Lifetime costs and QALYs for each combination of the 100 imputations and 100 parameter sets were analysed using seemingly unrelated regression adjusting for baseline values of risk factors in UKPDS-OM2.2 and Rubin’s rules [23, 29]. Variables were taken into account as a confounder, if the estimated regression coefficient for the cost or effect differences changed by 10% or more [24]. Cost-effectiveness acceptability curves for values of the willingness to pay ranging from 0 to 80,000 euro per QALY were constructed by assessing the fraction of the 10,000 combinations of imputation and parameter sets for which the NB of the FMD group was higher than the control group.
Since there is no information on the long-term adherence and effects of following the FMD, two different scenarios in the post-trial analysis were used. In the base-case scenario participants stop using the FMD directly after the end of the trial and risk factor progression follows the build-in equations of the UKPDS-OM2.2 model. In the alternative scenario, FMD participants who were still adhering to the FMD at the end of follow-up continue to use the FMD programme until the end of Life, but only 4 FMD cycles per year are used. The assumption was made that the change in weight, HbA1c and HDL-cholesterol observed in the FIT trial would similarly change in the same direction for one more year in these participants. These risk factors were chosen, since in the FIT trial a statistically significant difference between baseline and twelve months follow-up was found for these outcomes [15]. After this one year onwards, the assumption was made that risk factors would follow the build-in equations of the UKPDS-OM2.2 model. For participants who were not adhering to the FMD at the end of follow-up after one year, risk factor progression follows the UKPDS-OM2.2 model directly from the end of the trial. As sensitivity analyses for both scenarios, it was postulated that the price of the FMD programme is 50% lower due to frequent utilisation as an approved treatment. For a detailed description of the scenarios and sensitivity analyses, see Additional Methods in Supplementary Material.