Balancing Earlier Access and Data Immaturity – Are New Methods the Key?
In order to provide faster patient access to innovative oncology products, regulatory bodies are increasingly implementing accelerated approval schemes based on interim analysis or surrogate outcomes. A recent analysis of EMA approvals showed that 39% of approvals for oncology treatments from 2014–2017 were based on immature overall survival (OS) data, compared with only 19% from 2009–2013.1 Accordingly, manufacturers are submitting to HTA bodies “earlier”, with the resulting evidence submissions relying on clinical trial data that are limited by short follow-up and lack of power to measure differences in OS; 41% of “NICE” cancer technology appraisals published between 2015–2017 used immature data to inform reimbursement decisions.2 Given the immaturity of OS data, it’s no surprise that survival extrapolations are a key source of scrutiny and criticism, with extrapolated survival functions criticised by Evidence Review Groups (ERGs) in 71% of NICE cancer technology appraisals from 2011–2017.3
At this year’s ISPOR Europe, a number of sessions focussed on advances in methods for extrapolation and validation of long-term survival. Spotlight session 313 introduced two methods for leveraging external data or expert opinion: indirectly (to inform and validate model selection) or directly (to formally incorporate in extrapolation). Historically, extrapolation methods submitted in HTA appraisals have relied on indirect methods, for example, asking clinical experts to retrospectively assess the plausibility of extrapolations. However, the NICE Decision Support Unit (DSU) highlight in Technical Support Document 21 (TSD21) that such approaches are inherently subjective and thus may be prone to personal bias;4 my own judgemental bias was confirmed in Workshop 103 through participation in a real-time elicitation survey. Workshop 103 presented a new direct method for leveraging expert opinion, involving formal prior elicitation using a data book (a summary of relevant mortality data to provide context for estimates of survival), an elicitation survey involving training on risk of bias, and then direct incorporation of expert judgements into modelling via a Bayesian framework as informative priors (Figure 1).
Spotlight Session 313 presented a new flexible approach where two survival curves (an extrapolated curve based on the immature clinical trial data and another based on external registry data or an artificial dataset informed by external opinion) are “blended” in the long term (Figure 2).
These new methods may help improve the reliability of extrapolations based on short-term clinical trial data; however, they rely on the availability of existing external data from registries or expert knowledge about the expected survival profile in the long term. This may particularly limit their use in rarer indications with fewer patients, and hence more concentrated clinical expertise that is based on fewer case studies and less chance of reliable registry data. Challenges also remain in more common indications when evaluating novel, first-in-class treatments, where no such data or knowledge exists; in my experience, clinical experts are uncomfortable providing long-term survival estimates in these circumstances, so immature survival data from clinical trials may be the only available evidence to inform long-term extrapolations.
In such cases, how can HTA bodies have confidence in extrapolations, particularly if they are predicting survival benefits that have historically never been seen in that patient population? There is no easy answer to this question, but our research suggests that evidence from other indications or for similar technologies could help inform model selection. For a variety of innovative technologies, we retrospectively analysed the accuracy of OS extrapolations from interim data cuts by comparing predicted life years with realised life years calculated from long-term data cuts. At this year’s ISPOR Europe, I presented a case study investigating the accuracy of extrapolations for CAR-T therapy, which suggested that mixture cure models may be more appropriate than standard parametric models, aligning with the clinical plausibility of functional cure associated with these novel treatments.5 Previously, we also found that models reflecting non-monotonic hazards were consistently associated with more accurate predictions of long-term survival across indications for nivolumab, an immuno-oncology therapy.6
The challenges associated with immature data for the evaluation of oncology products aren’t going anywhere. Ensuring timely patient access to novel treatments will be incumbent on industry staying up-to-date with, and contributing to, the development of these new methods to improve the reliability of clinical input and reduce uncertainty in extrapolations, as well as HTA bodies remaining open to new methods and forms of evidence. We are yet to see widespread use of formal or direct elicitation in manufacturer submissions; perhaps we will start to see a determined shift towards these approaches in 2022.