A Comparison of Three Survival Models to Estimate The Cost-Effectiveness of Cancer Immunotherapy In The Treatment of Advanced Melanoma

Objectives:
To compare methods for extrapolating survival curves for previously-untreated patients receiving nivolumab versus ipilimumab for BRAF wild-type advanced melanoma (AM, comprising unresectable Stage III and/or Stage IV metastatic melanoma).

Methods:
Patient-level data from 203 patients with AM receiving nivolumab in a phase III study (nivolumab versus dacarbazine) were used to estimate hazards of progression and death. Of these, 56.2% (n=114) progressed and 23.2% (n=47) died during the study period. Weibull, log-logistic and a Weibull mixture cure model (MCM) were fitted to extrapolate trial data for overall survival (OS) and progression-free survival (PFS) up to a 10 year time horizon. To estimate transition probabilities for subjects receiving ipilimumab, hazard ratios were calculated by indirect comparison of nivolumab versus ipilimumab and applied to underlying survival distributions. Models were evaluated graphically, using Akaike’s Information Criterion (AIC) and naive comparison of the extrapolated ipilimumab survival functions with published long-term survival data.

Results:
AIC scores for the Weibull, log-logistic and MCM were 336.47, 335.30 and 776.20 for OS, respectively. The equivalent AIC scores for PFS were 511.39, 479.38 and 1421.63. The estimated 3-year survival rate of patients receiving ipilimumab was 2.8%, 15.1% and 55.8% and 14.9%, 40.2% and 87.2% for patients receiving nivolumab using the Weibull, log-logistic and MCM, respectively. This compares with published data showing approximately 21% (95% CI: 17-24%) of AM patients receiving ipilimumab (3 mg/kg) survive three years. Due to the short follow-up period for this study, the MCM introduced the greatest potential for error.

Conclusions:
The log-logistic model provided the closest approximation to real-world ipilimumab data. The choice of parametric model exerts a large effect on predicted effectiveness and cost-effectiveness of new therapies and needs justification. Different models should be considered in sensitivity analyses to estimate their impact on ICERs.