FormalPara Key Points for Decision Makers

The National Institute for Health and Care Excellence recently approved cannabidiol for reimbursement in the National Health Service of England for patients with tuberous sclerosis complex-associated seizures based on an average dosage of 15 mg/kg/day.

As current clinical experience with cannabidiol is better reflected by an average dosage of 12 mg/kg/day, the present study provides cost-effectiveness data at this dose to inform decision making.

1 Introduction

Tuberous sclerosis complex (TSC) is a rare (orphan) multisystemic disorder with a prevalence of 1 in 25,000 to 1 in 11,300 in Europe [1, 2]. Characterized by the formation of benign tumors in multiple organ systems [1], TSC commonly manifests with severe, debilitating neurological disorders, including epilepsy—prevalent in ~80% of patients [3,4,5].

TSC-associated epilepsy usually begins with focal seizures or infantile spasms [6], but can present with other, often treatment-refractory, seizure types, including generalized, myoclonic, and absence seizures [7, 8]. The characteristic infantile spasms are a known risk factor for both refractory epilepsy [7] and status epilepticus [9] that contribute to the increased morbidity [7, 9] and mortality in TSC [10].

In addition to epilepsy, patients often experience TSC-associated neuropsychiatric disorders (TAND), resulting in a complex burden of behavioral, psychiatric, intellectual, academic, and psychosocial disorders [11]. Together, TSC-associated epilepsy and TAND contribute to the impaired quality of life (QoL) of patients and caregivers [12, 13].

Although the pathogenetic mechanisms of TAND are unclear [14], uncontrolled seizures are a contributor to the development of these disorders and negatively impact QoL. For example, early seizure onset, infantile spasms, and treatment-refractory seizures are associated with poor cognitive outcomes/intellectual disability [15,16,17,18,19,20]. In comparison, early and effective seizure control is associated with improvements in cognition and QoL [21, 22].

Seizure control is therefore important, not just for patients and their caregivers, but for healthcare providers that deliver care within budgetary constraints. As a result of this complex burden of disease [10], medical costs in TSC can be three-fold higher than for an average member of the UK general population [23]. Key contributors to this cost include treatment-refractory seizures and episodes of status epilepticus that contribute to polypharmacy and increased healthcare resource utilization (HCRU; e.g. hospitalization and/or use of rescue medication) [5, 9, 24].

While the frequency of seizures is used to assess the burden of epilepsy and effectiveness of treatment, ‘seizure-free days’ is an important metric for patients and caregivers [25]. Seizure remission is not possible for many patients with TSC [26], and breakthrough seizures limit the potential QoL benefits associated with seizure remission. However, as seizures vary from day to day, there may be instances where treatment results in prolonged periods of seizure-free time that afford patients with periods of normality and lessen the impact of seizures on learning and memory [27]. Indeed, both patients and caregivers value seizure frequency and seizure-free time differently, with more seizure-free days valued as having a greater impact on QoL than fewer seizures [25]. Despite the diversity of antiseizure medications (ASMs) and non-pharmacological interventions, >60% of patients with TSC-associated epilepsy experience uncontrolled seizures and are at elevated risk of hospitalization and death [7, 23, 28]. For this high-risk group, there is an urgent need for alternative therapies. In the past 5 years, everolimus and cannabidiol have become available for the treatment of TSC-associated seizures—everolimus as a later-line treatment for focal-onset seizures, with or without secondary generalization [29], and cannabidiol for TSC-associated seizures, which may include focal and generalized seizures [30].

In the GWPCARE6 randomized placebo-controlled trial (ClinicalTrials.gov identifier: NCT02544763), the efficacy and safety of cannabidiol was investigated in patients ≥ 1 year of age with treatment-refractory TSC-associated seizures [31]. In the trial, TSC-associated seizures decreased by 48.6% with cannabidiol (25 mg/kg/day), a significantly greater reduction than observed with placebo (26.5%; p = 0.001) [31]. Communication of these results, and the acceptable safety profile, led to the approval of highly purified cannabidiol (Epidiolex® in the USA and Epidyolex® in the UK and EU [GW Pharma [International] B.V., now part of Jazz Pharmaceuticals, Inc.]) as an adjunctive treatment of seizures associated with TSC in patients ≥ 1 year of age in the USA and ≥ 2 years of age in the UK and in the EU [30, 32]. In line with the UK label, the National Institute for Health and Care Excellence (NICE) recently approved cannabidiol for reimbursement within England for treatment of TSC-associated seizures [33]. Following the NICE appraisal of cannabidiol, communication of cannabidiol’s cost effectiveness in treating TSC-associated seizures would be beneficial for patients, caregivers, and clinicians to inform treatment decisions.

The objective of the present study was to evaluate the cost effectiveness of add-on cannabidiol, compared with usual care, for the treatment of patients aged ≥ 2 years with TSC-associated seizures.

2 Methods

A cohort model was developed in Microsoft Excel® from the perspective of the National Health Service (NHS) in England and Wales to determine the impact of cannabidiol plus usual care and usual care alone on health-related quality of life (HRQL) and HCRU. This model also utilized a novel approach, using mixed-effects regression models in a cohort model for the evaluation of treatments in TSC-associated epilepsy. This approach was preferred over alternatives, such as a response-based modeling approach or a Markov state-transition approach, as it does not require the pre-specified division of the patient population into health states and instead allows the modeling of outcomes on a continuous scale. In addition, it allows prediction of the relative effects of treatment over time within a single model using covariates. Consistent with previous economic analyses of cannabidiol [34, 35], a lifetime horizon was modeled to capture costs and outcomes over a sufficient period, reflecting the lifelong nature of the disease and potential for adult-onset/treatment-resistant seizures [6]. Based on the mean age in the GWPCARE6 intention-to-treat patient population (13.9 years), the lifetime horizon was 86 years. Costs and quality-adjusted life-years (QALYs) were discounted at a rate of 3.5%, in line with the NICE reference case [36]. The health economic analysis plan and model used in the study are available upon request and are in line with the NICE appraisal with the exception of the base-case cannabidiol dosage of 12 mg/kg/day (15 mg/kg/day in the NICE submission).

2.1 Patient Population

All patients enrolled into GWPCARE6 [31] were included based on a clinical diagnosis of TSC; documented history of epilepsy incompletely controlled by current ASMs; age 1–65 years; treatment with ≥ 1 ASM; and ≥ 1 TSC-associated seizure in ≥ 3 of the past 4 weeks. Patients in GWPCARE6 were treated according to usual care and were randomized to add-on treatment with cannabidiol (Epidyolex®/Epidiolex®; 100 mg/mL oral solution) or placebo for 16 weeks. Usual care included management with ≥ 1 ASM excluding oral mechanistic target of rapamycin inhibitors or felbamate (if used in the prior year). Patient demographics, characteristics, and more details of usual care are described elsewhere, alongside descriptions of the trial setting/location [31].

2.2 Seizure Types

The cost-effectiveness analysis examined the impact of treatment in GWPCARE6 on focal seizures with impairment of awareness and generalized seizures (tonic-clonic, tonic, clonic, or atonic). Focal to bilateral tonic-clonic seizures were categorized as ‘generalized’ for the purposes of the model since any impact on HCRU and HRQL for patients and caregivers is likely the same regardless of onset. Focal motor seizures without impairment of awareness/consciousness were excluded as they are difficult to recognize and record [37].

2.3 Cannabidiol Dosage

The base-case analysis used an average cannabidiol dosage of 12 mg/kg/day. This dosage was selected to best reflect what is used in real-world clinical practice while also taking into consideration the expected effectiveness at this dosage based on clinical trial efficacy data collected in the three approved indications for Epidyolex. In UK clinical practice, physicians prescribe cannabidiol for TSC-associated seizures based on the guidance in the summary of product characteristics (SmPC) [30]. Regarding dosage and dose titration, the SmPC states that the recommended starting dosage of cannabidiol is 5 mg/kg/day for 1 week, and, in the second week, the dosage can be increased to the maintenance dosage of 10 mg/kg/day [30]. Based on individual clinical response and tolerability, the clinician can consider increasing the dosage further, in weekly 5 mg/kg/day increments, up to the label maximum of 25 mg/kg/day [30]. With UK and European prescribers following this guidance, we sought to determine what is used in clinical practice to inform the model dosage, and this evidence was provided to NICE for the technology appraisal (TA) of cannabidiol [33, 38]. According to real-world findings from a retrospective German claims database, the dosage of cannabidiol used in patients with probable TSC was ~12 mg/kg/day in children and ~8 mg/kg/day in adults [33, 38, 39]. In addition, as part of the aforementioned TA, clinical experts experienced in the management of TSC in Scotland and Wales agreed that most patients would have a dosage of ~12 mg/kg/day [33]. As efficacy was not determined for cannabidiol in the GWPCARE6 trial and open-label extension (OLE) below 25 mg/kg/day [31, 40], further evidence was required to determine the suitability of GWPCARE6 data in the modeling. The primary evidence provided to NICE in the TA was a subgroup analysis of GWPCARE6 OLE data comparing the percentage change in the frequency of TSC-associated seizures at a dosage of ≤ 15, > 15−25, and > 25 mg/kg/day [38]. Across all timepoints in the study, seizure reductions were not significantly different in patients on ≤ 15 mg/kg/day versus those on a higher dosage—thus supporting the lower dosage used in the model.

These findings are also supported by a post-hoc analysis of GWPCARE6 data which demonstrated that efficacy of cannabidiol is established by Day 10 of treatment when the dosage can be anticipated to be < 25 mg/kg/day [41]. This finding was supported by the clinical experts in the TA, who stated that because cannabidiol is titrated slowly in clinical practice, a response can be established at a lower dosage than the 25 mg/kg/day used in GWPCARE6 [33]. Lastly, the phase III randomized controlled trials in Lennox-Gastaut syndrome (LGS) and Dravet syndrome (DS) (GWPCARE3 and GWPCARE2) demonstrated similar efficacy of cannabidiol at 10 and 20 mg/kg/day when comparing seizure reductions and proportions of responders (≥ 50% seizure reduction) [42, 43]. Based on these findings, 12 mg/kg/day was used in the base-case model for the previous economic analyses in LGS and DS [34, 35].

In summary, with the available evidence on the prescribing habits and effectiveness of cannabidiol in patients with TSC, and considering the average dosage used in previous economic models of cannabidiol, a dosage of 12 mg/kg/day was chosen for the present cost-utility analysis.

2.4 Model Overview and Structure

As limited data are available in TSC to inform the model concept, structure and key assumptions, a Health Technology Assessment advisory board, including health economists and clinical experts, was conducted. Based on expert guidance, a cohort-level cost-effectiveness model was developed to determine average costs and QALYs (per patient) associated with changes in seizure frequency and seizure-free days for patients treated with cannabidiol or usual care. Additionally, caregiver disutility was included in the model. Three health states were modeled: (1) randomized treatment (cannabidiol or usual care); (2) discontinued treatment (usual care only); and (3) death (general mortality, TSC-related and sudden unexpected death in epilepsy [SUDEP]), with a cycle length of 7 days (Fig. 1).

Fig. 1
figure 1

Model structure diagram. Solid lines represent the transition between health states permitted by the model. Circular lines represent patients remaining in the health state for the alive health states (blue boxes) and following mortality (gray box). Dotted lines represent the modeled effect of treatment on TAND outcomes in early childhood based on changes in seizure frequency. aTAND outcomes included intellectual disability, delayed development, behavioral issues, autism spectrum disorder, attention deficit hyperactivity disorder, and anxiety disorders. bCohort population distributions based on IPD from the 16-week treatment period of GWPCARE6. cDiscontinuation rates were applied based on the 16-week GWPCARE6 and 72-week OLE study. A long-term discontinuation rate was applied after the OLE period to capture the impact of non-response and other long-term factors. A stop** rule was applied every 6 months for 2 years to reflect discontinuation resulting from non-response over time. CBD cannabidiol, IPD individual patient-level data, OLE open-label extension, SUDEP sudden unexpected death in epilepsy, TAND TSC-associated neuropsychiatric disorders, TSC tuberous sclerosis complex, UC usual care

Full size image

The model included discontinuation rates to reflect patients stop** treatment during the 16-week GWPCARE6 treatment period [31] and the 72-week OLE [40]. A long-term discontinuation rate was applied after the OLE period to reflect patients discontinuing treatment due to non-response and other factors [40].

A stop** rule was applied in the model every 6 months for 2 years to reflect treatment discontinuation due to non-response. Consistent with the NICE recommendations for cannabidiol in LGS, DS, and the more recent TA in TSC, non-response to cannabidiol treatment, and subsequent discontinuation, was determined when a patient’s seizure frequency did not decrease by at least 30% compared with the 6 months before starting treatment [33,34,35]. In the model, the population who discontinued or stopped cannabidiol treatment would match the ‘alive and on treatment with usual care only’ health state at the time points of discontinuation and follow the usual care trajectory with associated costs and QALYs for the remainder of the model time horizon. The proportion of patients who would be expected to stop treatment at 6 and 12 months was based on the GWPCARE6 treatment period and the OLE study (72-week follow up). The proportions of patients who would be expected to stop treatment at 18 and 24 months were assumed to be the same as the proportion expected to stop at 12 months, as the OLE did not have sufficient follow-up data to calculate rates at 18 and 24 months.

To capture the known benefits of early effective seizure control on TAND outcomes [21], the effect of treatment on TAND outcomes was modeled for each alive health state based on changes in seizure frequency in early childhood. A two-round Delphi panel involving 10 clinical experts was conducted to elicit information on the relationship between TSC and TAND [44, 45]. Further details on the Delphi panel, modeled TAND components, and the relationship with treatment of seizures are described in Online Resource 1 and Online Resource Table 1 (see electronic supplementary material [ESM]).

The prevalence of each TAND aspect in the model was sourced from a natural history retrospective study of patients with TSC-associated epilepsy from the TuberOus SClerosis registry to increase disease Awareness (TOSCA) database [11]. The prevalence of TAND aspects reported was matched to the age bands used in the model analysis for the calculation of drug costs.

For both alive health states (randomized and discontinued treatment), mortality was accounted for using age-specific general population mortality rates [46]. Background TSC-related and SUDEP mortality were included at the same rate for cannabidiol and usual care [10, 47].

2.5 Health States Categorization

Each alive health state was categorized into three sub-health states to allow different HCRU and HRQL values to be applied (Fig. 2). Categories for HCRU and HRQL were derived during the Delphi panel survey [45] (see Online Resource 1 in the ESM) and a vignette study (involving consultation with clinicians and caregivers), respectively (see Online Resource 2, Online Resource Tables 2–4 and Online Resource Figs. 1–3 in the ESM) [48].

Fig. 2
figure 2

Model process diagram. aSeizure frequency per day categories are aligned to HRQL data collected by seizure type. HCRU healthcare resource utilization, HRQL health-related quality of life, IPD individual patient-level data

Full size image

2.6 Regression Models

As a result of low patient numbers and analysis of individual seizure types, the use of transition matrices in the cost-effectiveness model would likely result in low transition counts for some health states, and some transitions would not be observed at all, which could bias the model results. To account for this, mixed-effects regression models were applied sequentially to individual patient-level data from GWPCARE6 to predict seizure outcomes for cannabidiol and usual care per 7-day model cycle; regression models were fitted using the lme4 package in R [49, 50]. Binomial and negative binomial regression were used to estimate the proportion of seizure-free days per week and seizure frequency on seizure days, respectively. These models were selected based on their suitability for binary outcomes and small sample sizes with over-dispersed count data [51]; the use of negative binomial regression is also consistent with the analysis performed on the primary endpoint in GWPCARE6. This approach provides the flexibility to model seizure-free days and seizure frequency outcomes separately, which allows separate utility values to be applied to each outcome.

In order to define a patient as ‘seizure free over 7 days’ in the model, a cut-off of 6.61 days was applied. This was used because the binomial regression in the model predicts values on a continuous scale rather than predicting discrete integer values. Within the context of a binomial logistic regression, it is not possible to predict exactly 0 or 1, which would correspond to 0 days or 7 seizure-free days, respectively. This is due to the prediction asymptote to the value of 0 or 1. Therefore, a cut-off of 6.61 days was assumed based on the maximum predicted number of seizure-free days by the binomial regression model (6.61).

Random effects terms were used within the models to account for multiple observations from the same patient. Two levels of random effect were applied in the model: 1) random intercept—each patient may have a different intercept value and 2) random slope—each patient may have a unique rate of change in outcomes (i.e. some patients may improve faster than others while others may decline over time). An offset term to account for the differing ‘exposure’ was included in the negative binomial model to account for days where patients had no seizures that were excluded from the dataset (as the number of seizures on seizure days were modeled) and any missing days.

Model covariates included

  • Treatment (cannabidiol plus usual care or usual care alone)

  • Treatment cycle: a covariate for each 7-day cycle was included in the regression and used as a proxy for time in the analysis. This was because the change in seizure frequency and seizure-free days demonstrated with treatment during GWPCARE6 indicated that patient outcomes improve after more time on treatment [52]. As this is not expected to happen in clinical practice, a log transformation was used, which slows the improvement of outcomes over time

  • Treatment and treatment-by-cycle interaction: to capture the treatment effect of cannabidiol plus usual care compared with usual care alone over time

  • Average baseline seizure frequency per week (continuous covariate): while not a treatment modifier in the pre-specified subgroup analyses, this covariate was included to allow the range of seizures within and between patients to be explicitly modeled to accurately capture the patient population. It was assumed that ongoing seizure frequency depends on the number of seizures the patient experienced at baseline. To provide one overall estimate of the baseline seizure rate for both the regression models, the baseline seizure frequency covariate included days where patients had no seizures in the estimate

Validation of the regression models was provided by the good fit of the seizure-free days (Fig. 3) and seizure frequency data from GWPCARE6 (Fig. 4). Extrapolated seizure frequency data from the fitted seizure frequency model (Fig. 5) showed a reasonable fit to the observed OLE data; however, seizure-free days data were unavailable for the OLE period and could not be validated. Beyond the GWPCARE6 trial period, the relative treatment effect of cannabidiol was assumed to remain consistent over the model time horizon. This assumption was based on the consistent treatment effect observed for cannabidiol in OLE studies for TSC-associated seizures [40], DS [53], and LGS [54], as well as the US-based expanded access program including patients with TSC-associated epilepsy [55].

Fig. 3
figure 3

Observed and estimated seizure-free days during GWPCARE6. The relationship between observed seizure-free day data, during the 16-week treatment period of GWPCARE6, and predictions from binomial regression are shown. CBD cannabidiol, UC usual care

Full size image
Fig. 4
figure 4

Observed and estimated seizure frequency during GWPCARE6. The relationship between the observed seizure frequency data, during the 16-week treatment period of GWPCARE6, and predictions from negative binomial regression are shown. CBD cannabidiol, UC usual care

Full size image
Fig. 5
figure 5

Observed and estimated seizure frequency during the GWPCARE6 OLE. Dashed line indicates the start of the OLE. Data are presented for the cannabidiol 25-mg/kg/day arm from the core trial. 95% confidence intervals are presented for the observed data. The relationship between the observed seizure frequency data, during the OLE period of GWPCARE6, and predictions from negative binomial regression are shown. OLE open-label extension

Full size image

2.7 Drug Cost Inputs

The model included costs for cannabidiol and usual care. Subsequent treatment consisted of concomitant ASMs, which were assumed to be consistent with usual care. The base-case analysis used an average cannabidiol dosage of 12 mg/kg/day. A confidential patient access scheme price (matching the NICE TA) for cannabidiol was used in the analysis. Concomitant ASM treatments were applied equally to both arms and for the entire time horizon. ASM costs in 2022 were sourced from the Monthly Index of Medical Specialties and the electronic Market Information Tool [56, 57]. As multiple doses and formulations were available for each ASM, NHS prescription cost data were used to estimate a weighted average [58].

Treatment acquisition costs included a representative sample of ASMs used in clinical practice. To avoid the model becoming prohibitively complex, the number of ASMs used in the analysis was limited; of the 29 separate ASMs used in the GWPCARE6 study (GW Research Ltd. Protocol GWEP1521 Clinical study report: unpublished data), those with a minimum 10% usage in either pediatric or adult patients at baseline were included in the model, giving a total of 10 ASMs for usual care (Online Resource 3, Online Resource Table 5, see ESM). In addition to these ASMs, everolimus was included as an option later in the treatment pathway for patients with TSC-associated epilepsy despite it not being commonly used alongside cannabidiol in GWPCARE6 [31]. To estimate the prevalence of everolimus use in clinical practice, we used data from the TOSCA database. In this retrospective study of patients with TSC-associated epilepsy, the use of everolimus (for the treatment of patients with focal-onset seizures) in clinical practice was low, at approximately 7.7% [6]. In practice, the low use of everolimus is likely because of its high cost [59] (as it is mostly used as an immunosuppressant/oncology treatment) and its restricted indication in the UK for patients who have treatment-resistant TSC-associated seizures and have already undergone, or have been considered for, surgery and/or vagus nerve stimulation [60]. Therefore, in the analysis, a proportion of the cohort based on the TOSCA data (7.7%) received everolimus in the cannabidiol arm following treatment discontinuation and in the placebo arm at 2 years following the trial period.

Dosages for all ASMs and cannabidiol can vary by age and weight (Online Resource 3, Online Resource Tables 6 and 7, see ESM). Therefore, to account for differing doses, age bands (2–6, 7–11, 12–17, and ≥ 18 years) were used to calculate drug costs. To accurately account for the variation in patient weights, the average per-cycle treatment cost for ASMs was based on the age-group weights and distribution from GWPCARE6.

2.8 Healthcare Resource Utilization

HCRU data were sourced using the Delphi panel [45] (see Online Resource 1 in the ESM). The model included the costs for HCRU, including primary care visits, outpatient visits, hospitalization, additional support (e.g. educational—children only), and institutionalization (residential care—adults only). Direct HCRU unit costs were obtained from the Personal Social Services Research Unit (PSSRU) in 2021 [61] and the NHS reference cost schedule 2020–2021 [62], detailed in Online Resource 3, Online Resource Table 8 (see ESM).

Additional support and institutionalization (e.g. residential care) costs were sourced from the PSSRU 2021 [61]. Cost inputs for educational support were inflated using healthcare indices published in the PSSRU 2021 (i.e. cost year 2021) [61].

HCRU differs among patients depending on seizure frequency, with a high frequency associated with higher HCRU (as these patients may require more care, including increased Accident & Emergency visits and hospitalizations [63]). HCRU also differs by seizure type, with higher HCRU associated with generalized seizures. Seizure frequency categories informed by the GWPCARE6 trial data were defined to collect HCRU. A summary of HCRU estimates for pediatric and adult patients and by seizure frequency per week is provided in Online Resource 3, Online Resource Table 9 (see ESM).

2.9 Adverse Event, Monitoring, and TAND Management Unit Costs

Costs for adverse events and monitoring were obtained from the PSSRU 2021 [61] and the NHS reference cost schedule 2020–2021 [62], whereas costs for managing TAND were sourced from a 2010 pan-European study of brain disorder costs [64] and were inflated to a cost year of 2021 (Online Resource 3, Online Resource Table 10, see ESM).

Adverse event (AE) costs and disutilities were applied to all serious, severe, and treatment-related AE rates observed in the 16-week GWPCARE6 trial. Serious AEs, classified as severe and treatment related, were defined as ‘require or prolong hospitalization’ or ‘other medically important’ and were assumed to require hospitalization [31]. Costs were assumed to account for the duration of the AE and associated treatment and were applied per cycle. An AE disutility of − 0.061 per month (adjusted to a disutility per cycle) was sourced from a study of epilepsy treatment-related side effects [65] and applied per model cycle to examine the impact of any AEs from treatment.

Monitoring costs associated with ASMs were assumed to be identical in both arms, with the exception of liver function with cannabidiol [30]—costs for 1 year of monitoring were applied to the cannabidiol arm.

As the analysis assumed that the progression of TAND is reduced rather than prevented entirely, the model applied a reduction in the cost of TAND management and a utility increment associated with improvement of TAND symptoms. These were applied for a period of 5 years to the proportion of patients aged ≤ 6 years who, after 6 months of treatment, had a 50% reduction in seizures compared with baseline. The decision to apply these model adjustments in patients aged ≤ 6 years was informed by the Delphi panel [44, 45]. A utility increment for each of the separate aspects of TAND mitigation was calculated from the literature, based on the difference between a patient having the disorder (mild intellectual disability [66], autism spectrum disorder [ASD] [67], symptomatic attention deficit hyperactivity disorder [ADHD] [68], and patients with anxiety disorders) [69] and not having it (intellectual disability, ASD) or having better control of the disorder (controlled ADHD, anxiety treatment non-responders).

2.10 Health State Utilities

To calculate the QALYs associated with treatment, a vignette study using time trade-off (TTO) methods, valued by the UK general public, was conducted to generate utility weights for patients and their caregivers [48]. Vignette descriptions were designed to reflect the HRQL profile of people with TSC-associated epilepsy and their caregivers, and the TTO weights used are published elsewhere [48]. This analysis assumed an average of 1.8 caregivers per patient (with the exception of institutionalized patients where 1 caregiver per patient was assumed), consistent with the preferred assumptions in the economic appraisal of cannabidiol in TSC [33]. Alternative approaches, including use of utility data from GWPCARE6 and utilities sourced from literature reviews, were investigated; however, these failed to provide suitable data for the model analysis (see Online Resource 3 in the ESM).

2.11 Outcomes

Total life-years, costs, and the number of QALYs for patients and caregivers were compared for cannabidiol plus usual care and usual care alone. The cost effectiveness of cannabidiol was measured as an incremental cost-effectiveness ratio (ICER). A willingness-to-pay (WTP) threshold of £20,000–£30,000 per QALY gained was used to assess the cost effectiveness of cannabidiol plus usual care versus usual care alone. A disease severity modifier was included as a model scenario based on satisfaction of the updated 2022 NICE criteria considering the severity of disease within decision making [36]. The QALY weight applied was 1.2, based on an absolute shortfall of 12–18 QALYs.

2.12 Sensitivity and Scenario Analyses

The uncertainty of parameters and the effect on the model results were assessed using a one-way sensitivity analysis (OWSA) and probabilistic sensitivity analysis (PSA). Uncertainty around estimates was set according to the variance information of each source. When variance data were unavailable, the standard error was set to 10% of the mean (further details are in Online Resource 3, Online Resource Table 9, see ESM). To test the robustness of the base-case analysis, scenario analyses were conducted for important structural and methodological assumptions, including time horizon, stop** rule rates, TAND response population, number of caregivers, cannabidiol dosage, and inclusion of social and educational costs.

3 Results

3.1 Base Case

In the base-case analysis over a lifetime horizon, compared with usual care alone, cannabidiol plus usual care was associated with an ICER of £23,797. The base-case results, with consideration of the disease severity modifier, reduced the ICER to £19,831 (Table 1). Our analysis suggests that seizure-free days are the biggest driver of the incremental QALY gains associated with cannabidiol versus usual care alone and amounted to 78% of the increase in the incremental QALY gains per patient.

Table 1 Base-case analysis and key scenario analyses
Full size table

3.2 One-Way Sensitivity Analysis

The results of the OWSA demonstrate that the model is robust to changes in key parameters (Fig. 6). The ICER was most sensitive to variation of the stop** rule assessment rate applied at 6 months for patients with a seizure frequency ≥ 7 seizures per week. For this population of patients within the highest seizure frequency category, the ICER values ranged from £19,811 to £28,529. Other key drivers of the model included the patient utility values applied to seizure-free patients and the response rates used to estimate the proportion of patients who benefit from a reduction in TAND symptoms.

Fig. 6
figure 6

Top 10 parameters influencing the model ICER from a one-way sensitivity analysis. The uncertainty of parameters and the effect on the cost-effectiveness model results were assessed via a one-way sensitivity analysis. The top 10 parameters that created the widest range in the model ICER results are shown. CBD cannabidiol, HS health state, ICER incremental cost-effectiveness ratio, OLE open-label extension, pw per week, TAND tuberous sclerosis complex-associated neuropsychiatric disorders

Full size image

3.3 Probabilistic Sensitivity Analysis

The results of the PSA showed that compared with usual care alone, cannabidiol plus usual care was associated with an ICER of £27,761. The PSA results were broadly comparable with those of the deterministic analysis (Fig. 7); however, the incremental QALYs in the PSA and deterministic analysis diverge with increasing QALYs owing to categorization of continuous variables (for the purpose of estimating costs and QALYs) and inclusion of patients with high baseline seizure frequency in the PSA. At WTP thresholds of £20,000 and £30,000 per QALY gained, the probabilities of cannabidiol plus usual care being cost effective compared with usual care alone were 30% and 52% for the base case (Fig. 8) and 39% and 66% following inclusion of the NICE disease severity modifier.

Fig. 7
figure 7

Probabilistic sensitivity analysis for the model ICER. Probabilistic sensitivity analysis (PSA) results are shown for CBD plus usual care versus usual care alone (gray diamonds). Mean data are shown for the PSA (blue square) and a deterministic analysis (red square) alongside trendlines for the NICE ICER thresholds of £20,000 (red line) and £30,000 (blue line). CBD cannabidiol, ICER incremental cost-effectiveness ratio, NICE National Institute for Health and Care Excellence, QALY quality-adjusted life-year, UC usual care

Full size image
Fig. 8
figure 8

Probabilistic sensitivity analysis cost-effectiveness acceptability curve. The cost-effectiveness acceptability curve was generated using the results from numerous probabilistic model runs. The probabilistic sensitivity analysis was sampled from the distribution of each model parameter 1000 times. CBD cannabidiol, ICER incremental cost-effectiveness ratio, QALY quality-adjusted life-year, UC usual care

Full size image

3.4 Scenario Analysis

Scenario analysis results are shown in Table 1. To test the sensitivity of the model to patient HRQL, two scenarios were modeled assessing alternative sources of patient HRQL: one based on Tritton et al. (EuroQol 5 Dimensions, EQ-5D) [70], which resulted in an ICER of £22,153, and one based on Vergeer et al. (Health Utilities Index-3) [22], which resulted in an ICER of £26,177. In comparison, the most influential scenario was the inclusion of wider social care and educational costs, elicited via the Delphi panel study [45]. This scenario had the largest impact resulting in cost savings for the cannabidiol arm and a dominant ICER.

4 Discussion

Our cost-effectiveness analysis of cannabidiol provides a new approach for the evaluation of treatments in patients with TSC-associated seizures using regression models in a cohort model. Using this approach, our base-case model with a WTP threshold of £20,000–£30,000 provided evidence to aid decision making regarding the cost effectiveness of add-on cannabidiol in patients with TSC-associated seizures aged ≥ 2 years who are refractory to current treatment.

Based on a WTP threshold between £20,000 and £30,000 and modeled average dosage of 15 mg/kg/day, NICE recently approved cannabidiol for reimbursement in England for management of TSC-associated seizures [33]. As outlined in the methods, the lower than maximum label (< 25 mg/kg/day) dosage in the model was informed by real-world clinical practice whereby clinicians follow the label guidance and titrate to a dosage of 10 mg/kg/day and only titrate further following consideration of individual benefit and risk [30]. A subgroup analysis of GWPCARE6 OLE data supports the efficacy of cannabidiol at dosages < 25 mg/kg/day. These data demonstrated a lack of dose response for cannabidiol when comparing percentage seizure reductions from baseline at dosages of ≤ 15, > 15−25, and > 25 mg/kg/day [38]. While we calculated ICERs using a base-case cannabidiol dosage of 12 mg/kg/day, and the TA acknowledged this dosage is likely the average used in clinical practice, NICE considered 15 mg/kg/day a more conservative estimate [33]. Support for a dosage of 12 mg/kg/day comes from a retrospective analysis of German claims data which estimated the average cannabidiol dosage prescribed in patients with probable TSC to be < 13 mg/kg/day [39]. Use of 12 mg/kg/day in the base-case model is also consistent with previous economic analyses in LGS and DS [34, 35]. Interestingly, there is evidence that adults are prescribed lower cannabidiol doses than children, potentially reflecting age-related differences in drug metabolism [71, 72].

Several economic evaluations have been published investigating the cost effectiveness of cannabidiol in childhood-onset epilepsies with contrasting conclusions [73,74,75]. Briefly, two articles have found cannabidiol to be cost effective—one from the perspective of the Canadian healthcare system, which found cannabidiol to be cost effective versus usual care (ICER of Can$32,399 and WTP threshold of Can$50,000) over a 13-year time horizon in children with DS [74], and another which summarized the evidence submitted to NICE regarding the cost effectiveness of cannabidiol in LGS and DS (ICERs of £33,721 and £32,471, respectively) [73]. In contrast, cannabidiol was not considered cost effective, over a lifetime horizon, from the US payer perspective in children with LGS (ICER of US$451,800 and WTP threshold of US$150,000) [75]. Methodological differences in the study design and assumptions of the models likely underpin the differing conclusions, and these have been discussed previously [76, 77]. The present study is most closely aligned with the NICE TA in LGS and DS, finding cannabidiol to be cost effective, where patient-level data of seizure frequency and seizure-free days were modeled, and impacts on patient and caregiver QoL were incorporated into QALYs.

The economic model was developed to fully capture the burden of TSC-associated epilepsy and the potential benefits of treatment on the QoL of patients and caregivers. Firstly, the regression modeling approach used allowed the proportion of seizure-free days, and not just the seizure frequency, to be estimated on a continuous scale using all available data. Using covariate-adjusted models, it was then possible to predict the relative effects of treatment over time. Secondly, inclusion of TAND and caregiver utility in the model allowed for the broader benefits of treatment on patients and caregivers to be considered. Using this approach, we demonstrated that the increased costs associated with cannabidiol are offset by the decreased health state costs in the higher seizure frequency health states (> 2 to ≤ 7 seizures per week and > 7 seizures per week). In addition, for patients with a significantly lower lifespan and QoL versus the general population, NICE guidance allows application of a disease severity modifier to reflect the severity of disease and anticipated QALY shortfall [36]. Considering the severity of TSC-associated epilepsy on the QoL of individuals and their caregivers [12, 13], our analysis additionally demonstrated, following the application of a disease severity QALY modifier of 1.2, an ICER reduction of 16.7% compared with the base-case analysis. While we tried to fully capture the burden of TSC-associated epilepsy in our model, treatment with cannabidiol may result in additional positive impacts on patients and caregivers that were not modeled here (e.g. a reduction in the duration/severity of seizures, the long-term impact of improved seizure control on comorbidities and injuries, as well as the risk of SUDEP).

Our findings are consistent with the clinical benefits of cannabidiol on TSC-associated seizures [31] and the anticipated benefits on HCRU that follow when improved seizure control contributes to fewer hospitalizations, injuries, falls, and fractures. Evidence for this assumption has been demonstrated in a cohort of patients with treatment-refractory epilepsy, where cannabidiol decreased hospital admissions [78]. Further, given the potential for improved epilepsy control to prevent seizure-related brain damage [21] and cognitive decline [16, 21], it would be anticipated that cannabidiol could also improve the burden of TAND and the associated poorer QoL [12] when seizures are controlled early in TSC before the onset of significant developmental impairment.

The sensitivity analyses, adjusting for covariates including TAND population and social and educational resource use, showed that the results of the base-case analysis are robust, with the inclusion of educational resource use returning a dominant ICER. All other scenarios modeled demonstrated ICERs below the £30,000 WTP threshold, including two scenarios using published sources of patient utility. Additionally, our analysis suggests that seizure-free days are the biggest driver of incremental QALYs associated with cannabidiol plus usual care compared with usual care alone, equating to 78% of the incremental QALY gain per patient. This agrees with a previous study, in caregivers of patients with LGS and DS, that demonstrated additional seizure-free days improved patient QoL to a greater extent than fewer seizures [25]. Consistent with this, it has been suggested that seizure-free days, in addition to measurement of seizure frequency and responder rates, may be an important indicator of ASM effectiveness because of their potential to improve patient and caregiver QoL [79].

Regardless of impact on seizure-free days, ASMs can have a significant positive impact on patient outcomes, and depending on the healthcare resources available, may be the most cost-effective option. This was demonstrated by Fallah et al., with carbamazepine (as the third ASM) and clobazam (as the fourth ASM) in children with treatment-refractory TSC-associated seizures [80]. In contrast, everolimus may not be as cost effective as these ASMs over a 5-year period [80], which justifies the positioning of everolimus in our model as a later-line treatment applied equally to a small proportion of patients in both arms.

Given the rarity of TSC [1], there were data gaps that limited the cost-effectiveness model. To account for these, we used robust elicitation methods in the model analysis, including a two-round Delphi panel study to collect information on HCRU and TAND [44, 45]. In addition, numerous challenges are associated with collection of HRQL from patients with severe and treatment-refractory epilepsies like TSC-associated epilepsy. For example, there are no validated disease-specific instruments in this population [81], and generic QoL instruments (e.g. EQ-5D) do not strongly correlate with disease-specific measures for epilepsy [82]. To generate HRQL data for the model population, we conducted a vignette study in the general population using TTO methodology to elicit patient and caregiver HRQL, and the ICER was robust to scenario analysis using alternative sources of patient HRQL [22, 48, 70].

The model used the same 16-week treatment period as GWPCARE6 [31]. Based on the short-term nature of the trial, there is some uncertainty on the long-term effect of cannabidiol. However, the GWPCARE6 OLE data provide a direct validation of a sustained treatment effect for patients who remain on treatment over the long term [40]. Our finding that the extrapolated seizure frequency data from the fitted frequency model align with the observed OLE data suggests that the long-term predictions from the regression model are reasonable, and we can be confident regarding the antiseizure effects of cannabidiol modeled throughout the blinded trial and OLE period. A sustained clinical benefit is further supported by data in LGS and DS [53, 54].

Further limitations relate to the lack of data to determine HCRU in patients with TSC and the long-term impact of cannabidiol on TAND. To account for these limitations, we conducted a two-round Delphi panel study to derive hospital admission and social and residential HCRU data, as well as determine the potential positive outcomes associated with effective treatment of seizures in a pediatric TSC population [44, 45]. While there are merits to the Delphi panel approach, when published data are lacking to inform economic evaluations [83], the information is collected from a limited pool of experts (in this case 10 clinical experts), which may lead to bias that can be amplified by the approach used to determine consensus. However, as with any rare disease, there are a limited number of experts to inform Delphi panels, which can be considered as an additional limitation when used in the present setting. Lastly, data from the literature were used to inform the cost and QALY outcomes associated with TAND, and we explored several scenarios to address uncertainty and the limitations of this approach.

5 Conclusions

Based on an ICER threshold of £20,000–£30,000 for England and Wales, and an average dosage of 12 mg/kg/day, we provide evidence for the cost effectiveness of cannabidiol as an add-on treatment in patients with TSC. Probabilities of cost effectiveness for cannabidiol plus usual care compared with usual care alone at WTP thresholds of £20,000 and £30,000 were 30% and 52%, respectively, for the base case and 39% and 66%, respectively, following inclusion of the NICE disease severity modifier. These findings were robust to sensitivity and scenario analyses, validating the chosen model parameters to address important data gaps in patient and caregiver HRQL and HCRU. Cannabidiol should have a limited budget impact on the NHS due to the orphan nature of TSC-associated epilepsy and the cost offsets associated with disease management.