OBJECTIVES: Overall survival (OS) is an important metric when evaluating the benefits of immunotherapies such as nivolumab, but clinical trial data often requires extrapolation to predict long-term survival. For instance, in health technology assessments (HTAs), decisions must be made based on immature data in order to provide timely access to novel therapies. This study compared predicted OS values from extrapolation models informing two first-line gastro-oesophageal NICE HTAs (TA865 and TA857) with values based on updated data from their informing trials.

METHODS: To inform the HTAs, models selected by the company, evidence assessment group (EAG) or assessment committee (AC) were developed from immature patient-level data. Area under predictive survival curves (AUC) from these models were used to estimate accrued life-months to 36 months and compared to the AUC of Kaplan–Meier (KM) curves of OS from recent data cut-offs of informing trials.

RESULTS: In TA865, the company preferred a piecewise KM/lognormal model, whilst the EAG preferred a fully parametric log-logistic model. AUC was 12.5, 12.3 (company, EAG) versus 12.4 (KM) months for chemotherapy and 17.5, 17.6 versus 17.8 months for nivolumab + chemotherapy. In TA857, the company preferred a 4-state semi-Markov model (SMM) using a long-term response (LTR) state; the EAG removed the LTR state from the SMM, and the AC requested that the model be rebuilt as a 3-state partitioned survival model (PSM). AUC was 15.3, 15.2, 14.0 (company, EAG, AC) versus 13.9 (KM) months for chemotherapy and 17.8, 17.7, 17.7 versus 17.6 months for nivolumab + chemotherapy.

CONCLUSIONS: Continued follow-up validates predictions used in the models informing TA865 and TA857. In TA865, the company and EAG models show better fit to opposite arms of the trial, neutralising incremental differences in predictions. In TA857, all models predicted accurately for nivolumab + chemotherapy, whilst the PSM improved prediction for chemotherapy.