OBJECTIVES: Due to their time-budget constraints and the increasing efficacy of novel therapies, clinical trials often present immature overall survival (OS) data, and regulatory bodies commonly rely on extrapolation procedures to evaluate survival benefits beyond trial follow-up. This study investigates parametric models’ reliability in extrapolating long-term survival projections based on immature OS data.

METHODS: Phase 3 clinical trials of drugs reimbursed in Italy in the last three years (2021-2023) for any of the five cancers causing more deaths, having OS as primary endpoint and reporting mature Kaplan-Meier (KM) survival data were included in the study. Each OS was digitalized, and an immature version of the curve was derived by cutting mature dataset at 24 and 12 months respectively for first-line and subsequent line indications, then both immature and mature KM data were extrapolated over a 10-year horizon, using six parametric models. Each model’s fitness was evaluated by assessing the sum of the Akaike and Bayesian Information Criteria (AIC+BIC) and extrapolations of immature and mature data obtained using the same model were compared by calculating the mean squared error (MSE) between each point of the two curves. Models showing an overall MSE<0.0005 were considered reliable.

RESULTS: Ten clinical trials were included. The lognormal distribution emerged as the fittest for immature data in 4 cases, followed by the loglogistic distribution in 3 cases. In 8 cases the model with the lowest AIC+BIC value for immature data was also reliable, and in 3 occurrences it also resulted in the lowest MSE. Among these 8, fittest model for immature and mature data corresponded 7 times, in the mismatching case the model resulted second fittest.

CONCLUSIONS: This study showed that long-term extrapolations of immature OS data are generally reliable, especially when the fittest model for immature data is the lognormal or loglogistic distribution.