Cost-effectiveness of voretigene neparvovec in the treatment of patients with inherited retinal disease with RPE65 mutation in Switzerland

The model’s Markov structure considers five health states (HS) related to visual acuity (VA)/ visual field (VF), and a death state. The best vision-related health state patients could be in was 'moderate visual impairment' (i.e. VA>1 on LogMAR (Logarithm of the Minimum Angle of Resolution) scale as the average of both eyes and/or VF>240 degrees as the average of both eyes) and the worst was 'hand motion, light perception or no light perception' (i.e. VA≥3 on LogMAR scale as the average of both eyes, or indications of hand motion, light perception or no light perception across both eyes). Criteria used for categorising patients into a health state were derived from American Medical Association guidelines and are provided in appendix Table 1. Transitions between health states are presented in Fig. 1. During each model cycle, patients were simulated to either remain in the same health state, transition to a better (only possible in the first year) or worse vision-related health state, or die.

The population of interest were IRD patients with a biallelic RPE65-mutation and sufficient viable retinal cells. Patient characteristics were assumed to reflect those of the Study 301/302 (see below) trial population [8], with a mean age of 15.1 years at baseline, 23% of patients in HS1, 32% in HS2, 23% in HS3, 19% in HS4 and 3% in HS5.

The clinical effectiveness of VN treatment was based on Study 301/302 trial data (ClinicalTrials.gov: NCT00999609) [7, 8, 10]. Study 301/302 is the only Phase 3 randomised controlled trial of VN for IRD patients with a biallelic RPE65-mutation. The study found a statistically significant improvement in multi-luminance mobility testing (MLMT) performance (primary outcome) at 1-year follow-up (MLMT improvement=1·6; 95% CI: 0·72 to 2·41). The MLMT measures functional vision, by assessing the ability of an individual to navigate an obstacle course accurately and at a reasonable pace at different levels of illumination.

The Markov model consisted of two phases to encompass the lifetime horizon of the analysis: an initial phase for the first year, and a long-term phase for the subsequent 84 years. Transition probabilities for the initial phase were calculated by assessing VA/VF outcomes in the first year of the Study 301/302 trial, with a further assumption made for patients in HS5 at baseline (Table 1). This assumption was made because no Study 301/302 patients in the VN or SoC strategies who were in HS5 at baseline were observed at 12 months (due to participant withdrawal). Therefore, it was assumed that for the initial phase, transition probabilities for HS5 patients at baseline would move in the same direction and same number of steps as the HS4 patients in each treatment strategy (e.g. as 100% of SoC patients in HS4 at baseline were in HS3 at 12 months, it was assumed 100% of SoC patients in HS5 at baseline were in HS4 at 12 months).

After year 1, for VN patients the VN treatment effect was assumed to be retained for the next 39 years. This meant that apart from the VN patients projected to transition to death, for 39 years VN patients were assumed to remain in the same health state that they were in at the end of year 1. A treatment effect duration of 40 years was assumed as it represents the midpoint between a minimum of 7.5 years (based on available Phase 1 trial follow-up data for VN [24]), and a maximum of 70 years (assuming the treatment effect lasts for the entire lifetime of patients [23]). 40 years was accepted to be the most plausible treatment effect duration by the NICE committee for the UK VN submission [23]. Transition probabilities representing visual decline during the long-term phase were based on a Weibull survival regression derived from a retrospective dataset of 70 IRD patients with RPE65-mutation (Appendix Table 2) [25, 26]. The Weibull model was chosen as it exhibited best statistical fit (assessed by Akaike and Bayesian information criterions) out of the different distributions tested. The Weibull regression was applied after the 1st year of the model for SoC patients and applied after the 40th year for VN patients. The multistate Weibull regression model used provides estimates of the transition intensities (which are converted to transition probabilities) at each cycle in the economic model. This effectively means that in each cycle a different transition probability matrix is used. The transition probabilities are therefore time-dependent, however this ‘time’ is time from baseline (and not time since entering a health state). The assumption made is the classical Markovian assumption of memorylessness and therefore this does not require knowledge of time spent in each health states. All transition data was pooled to inform a single multistate survival model which allowed the calculation of transition probabilities in each cycle based on time from the start of the model. An overview of the use of multi-state models for the analysis of time-to-event data is provided by Meira-Machado et al [27]. A reference transition from HS1 to HS2 was used, and hazard ratios representing other transitions (e.g. from HS1 to HS3) applied to calculate the probability of making each transition. Mortality was modelled based on Swiss lifetables [28].

Health state utilities were derived from a study where EQ-5D-5L estimates for IRD patients with biallelic RPE65-mutation were elicited from clinicians (it was not considered feasible to collect EQ-5D-5L data directly from a representative sample of patients due to the rare nature of the disease) (Table 2) [5]. Utility decrements resulting from the adverse events in Study 301/302 were included for the adverse events which occurred in at least 2 patients, could be attributed to VN treatment/administration procedure, and are associated with a utility reduction [22]. These adverse events were cataracts, eye inflammation, and increased intraocular pressure (IOP), and were incorporated into the analysis in the form of a one-off QALY loss for VN patients (Table 3). The proportion of VN patients affected by these adverse events were taken from Study 301/302. The utility decrements of cataracts and eye inflammation and assumed duration of these adverse events, were obtained from a NICE report estimating these decrements for age-related macular degeneration patients [29]. Due to the absence of a study estimating the utility decrement of increased IOP, it was assumed to be the same as the utility decrement of uncontrolled/severe glaucoma which was obtained from a study by Pershing et al [30]. Duration of increased IOP was assumed to be one month as all increased IOP events in study 301/302 were fully resolved by 1 month.

In the base-case analysis, for the VN strategy health-care costs included eligibility testing, VN acquisition, administration of VN and treatment of adverse. For both treatment strategies, costs included those generated due to impaired vision, in the form of technical assistance including vision aids, community and residential care services for visually impaired patients over 65 years of age. All resource use parameters were taken from published studies, and – if required – verified to be applicable to Switzerland by Swiss clinical experts (Table 4). Unit costs for outpatient and inpatient services in Switzerland were derived from national sources [28, 31‐37] (Table 4). The planned public price of VN (Swiss francs (CHF) 759'968) was used in the analysis. In a complementary analysis, a societal perspective was adopted. In this, indirect costs in the form of productivity losses from impaired vision, in the workplace for patients and caregivers, as well as increased education costs for patients aged below 18 years were additionally included. These were estimated from Swiss national statistical data [31, 35, 36].

In order to assess the robustness of the results, a univariate sensitivity analysis and a probabilistic sensitivity analysis (PSA) of the base case results for the healthcare system perspective were performed in addition to a range of scenario analyses.

In the univariate sensitivity analyses, parameters were varied by their 95% confidence limits if available, and otherwise by ±20%. The following groups of parameters were included: Weibull regression parameters representing visual decline in the long-term phase, parameters representing health state utilities or adverse event related utility decrements, parameters representing duration and probability of AEs, resource use parameters, and cost parameters (excluding the acquisition cost of VN, which was treated as fixed). A Tornado diagram was produced; the presentation of results were restricted to the 10 most influential parameters.

A probabilistic sensitivity analysis using 10’000 iterations was performed. Normal distributions were used for the Weibull regression parameters representing visual decline in the long-term phase. Beta distributions were used for parameters representing probabilities and absolute utilities. Gamma distributions were used for parameters representing utility decrements, resource use and costs.

In addition, several scenario analyses were conducted. The univariate sensitivity analysis did not cover the transitions between health states occurring in the first year of the model [7, 8]. To account for uncertainty in treatment effectiveness, the assumption on the duration of the VN treatment effect was varied. In the base-case analysis the treatment effect was assumed to last for 40 years; in scenario analyses alternative durations of 7.5 years (which is the currently available Phase 1 trial follow-up data for VN), 20 years and life-long have been adopted. In another set of scenario analyses, alternative distributions have been used to model the visual decline of IRD patients with RPE65 mutation during the long-term phase of the model. Also, alternative health state utilities (Table 2) were implemented in a scenario analysis; using Health Utilities Index Mark 3 (HUI3) estimates obtained from the same clinicians from which EQ-5D-5L estimates were obtained for the base-case [5]. In another scenario, health states were defined according to visual function only (in the base-case, health states were defined according to the worst of visual acuity or visual function). In further scenarios, alternative discount rates of 0% to 5% were applied. In another scenario analysis, we assigned health states according to the best-seeing eye, rather than the average eye. In an additional scenario analysis, for health state transitions with no data (from HS5), we assume patients remain in the same health state instead of assuming the movement is the same as the previous state.

We carried out two complementary analyses adopting a societal perspective. In the first, we included productivity losses of patients, by assuming patients of working age with severe visual impairment are 80% less productive than those without. This assumption was made based on a study from India [45]. In the second, we more conservatively assumed patients are 50% less productive as estimated from a Royal National Institute of Blind People (RNIB) study in the UK [23].

The original model that was adapted, was validated using standard procedures including tests to check model outputs were logical, and cell-by-cell checks of consistency and logic [23]. The model was implemented in Microsoft Excel 2010.