OBJECTIVES: The Guyot method is widely considered as the ‘gold standard’ for reconstructing individual-level patient survival data (IPD) from digitised Kaplan-Meier curves. However, there are limitations and caveats to this approach. Although the Guyot method performs well for reproducing survival curves, it may have limited performance for reproducing (or estimating if not reported) other summary measures such as the hazard ratio (HR), especially when full information (risk table and total number of events) is unavailable. This limited performance is magnified when the proportional hazards assumption is violated. With the objective of estimating summary measures not reported with uncertainty, especially with crossing survival curves, a new ‘simulation approach’ was developed for reconstructing IPD.

METHODS: The simulation approach consists of 8 stages. In summary, we simulate survival times from a log-cumulative hazards model with restricted cubic splines and censoring times using a piecewise exponential distribution. Uncertainty in the reconstruction process is obtained by producing multiple datasets from the data generating model. We average over the results to obtain a final point estimate for the target summary measure. The simulation approach is compared with the Guyot method for estimating summary measures in the presence of non-proportional hazards (NPH).

RESULTS: We demonstrate improved performance with respect to bias compared to the original IPD estimate (with reported uncertainty) for estimating alternative summary measures to survival probabilities such as the (average) HR and restricted mean survival time (RMST) in the presence of NPH.

CONCLUSIONS: In the presence of NPH, reconstructing IPD using the Guyot method to obtain summary measures that are not reported may be inappropriate. Alternatively, a simulation approach is presented which shows good performance for reproducing IPD and estimating summary measures with uncertainty. This can be easily extended for correlated endpoints such as PFS and OS.