Reconstructing patient level survival data from published Kaplan-Meier curves

Individual-level patient data (IPD) are helpful for designing clinical trials, conducting meta-analyses, or methodology research. However, such patient level data are not readily available. Multiple methods have been developed for reconstructing survival data using published Kaplan-Meier (KM) survival curves. There has been no practical guidance on an optimal approach or extensive evaluation of the performance of the approach.

We reviewed several methods of extracting the coordinates of KM survival curves and reconstructing individual-level survival data. Then, we reproduced data from 46 published KM curves. The accuracy of reconstructed data is quantified by comparing hazard ratios (HRs) and their confidence intervals (CIs) estimated from the reproduced data with those reported in the original papers.

The comparison showed a high degree of similarity between the reproduced and original HRs and CIs. In most cases, the differences were less than 5 %. The mean and median absolute percentage differences of 58 reconstructed HRs were 2.85 % and 2.14 %, respectively. These results suggest the reconstruction method reliably reconstructs survival data from KM survival curves.

Based on an extensive number of reconstructions, we demonstrated that reconstructed data provided similar estimates overall to those from published papers. The quality of the reproduced data depends on the presence of noise in the published curves and whether the preprocessing step is properly done.