OBJECTIVES: Treatment waning is a common uncertainty in health technology assessment, and choice of waning methodology can have a substantial impact on estimated survival.¹ This research investigated the accuracy of four waning methods used in past National Institute for Health and Care Excellence appraisals in predicting overall survival (OS) for immuno-oncology therapies, using pembrolizumab in KEYNOTE-024 as a case study.

METHODS: Four waning methods were applied to pembrolizumab OS data from a 25-month data-cut of KEYNOTE-024. Methods 1 and 2 assumed full treatment effect until 5 years, after which all effect was lost relative to chemotherapy. Methods 3 and 4 linearly waned treatment effect between 2 and 5 years, based on the trial two-year stopping rule.¹ Predicted LYs were calculated over 5.5 and 10-year horizons, using best-fitting curves for pembrolizumab (Gompertz) and chemotherapy (lognormal). Predicted LYs were compared with more mature LY estimates from a later (5.5-year) data-cut of KEYNOTE-024, calculated over a 5.5-year time horizon (using pembrolizumab Kaplan-Meier data directly), and over a 10-year horizon (by extrapolating the Kaplan-Meier data using the best-fitting curve [log-logistic] with no waning applied).

RESULTS: Predicted LYs over a 5.5-year horizon ranged from 2.81 (Method 4) to 2.93 (Method 1). Over a 10-year horizon, predicted LYs ranged from 3.69 (Method 4) to 4.38 (Method 1). The more mature LY estimates were 2.74 and 3.87 over a 5.5 and 10-year horizon, respectively. LYs predicted by Method 4 aligned most closely with the more mature LY estimates when calculated over both time horizons.

CONCLUSIONS: Waning based on gradually equalizing hazards (Method 4) demonstrated the greatest predictive accuracy compared to the other methods explored. Further research is needed to confirm whether this finding is generalisable to other treatments and indications.