Web Tool Accurately Reconstructs Survival Data From Plots

Researchers have developed a web-based application called V-Exploder that allows investigators to reconstruct individual patient data from published Kaplan-Meier survival plots with high accuracy. The tool integrates the complete workflow of data extraction into a single browser interface, offering immediate visual feedback that enables more precise reconstruction than previously possible with established toolchains. Published in BMC Medical Research Methodology, the study describes the implementation of this academic, non-profit, data-local tool and reports on an evaluation study assessing its performance across various reconstruction methods and dataset sizes.

Individual patient-level survival data are essential for various types of medical research, including meta-analyses and cost-effectiveness analyses. However, such data from clinical trials are usually not publicly available, forcing secondary data utilization to rely on reconstructing data from published Kaplan-Meier plots. Existing toolchains for this reconstruction are often cumbersome and rarely offer possibilities to assess reconstruction accuracy during the process. V-Exploder addresses these limitations by allowing users to superimpose recreated graphs onto the original figure continuously, facilitating precise alignment and correction.

The application implements three distinct methods for reconstructing individual patient data. These include the iterative algorithms described by Guyot et al. and Rogula et al., as well as a novel direct method for the positioning of individual events and censorings. The direct method allows investigators to manually set individual events and censorings rather than specifying coordinates of steps in the graph, providing maximum control over the reconstruction process. This approach eliminates the need for an algorithm to generate data from survival estimates, a concept the authors note has not been previously described in the literature due to the requirement for a rich and dynamic user interface.

To evaluate the performance of V-Exploder, the research team conducted a study involving ten investigators who each reconstructed data from four of a total of twelve figures. Each figure showed the survival of two randomly generated cohorts with sizes of 20, 50, 100, or 200 subjects. Additionally, nine investigators reconstructed data from a publicly available real-world dataset from The Cancer Genome Atlas showing the survival of 200 patients with acute myeloid leukemia. The investigators had varying levels of experience, with most reporting having performed few or no prior reconstructions, and all received a single introduction to the software.

The study measured the similarity between reconstructed and original graphs by calculating the absolute area between the curves and performing survival regression to obtain regression parameters. Survival metrics commonly used in analysis, including survival quantiles, landmark survival, and log-transformed hazard ratios, were calculated for both the reconstructed and reference data. The researchers compared these metrics to assess validity and reliability, reporting median absolute errors, their interquartile ranges, and root mean square errors. This comprehensive approach allowed for a detailed comparison of the accuracy provided by each of the three implemented reconstruction methods.

Results indicated that survival metrics from reconstructed data showed high similarity to those calculated from the reference data across all datasets, investigators, and methods. Regarding log-transformed hazard ratios, the median absolute error with the direct method was 5.93 × 10⁻⁴, with an interquartile range of 7.61 × 10⁻⁵ to 1.65 × 10⁻³, and a root mean square error of 6.26 × 10⁻³. The Rogula method yielded a median absolute error of 5.28 × 10⁻⁴ and a root mean square error of 4.05 × 10⁻³. In contrast, the Guyot method showed a higher median absolute error of 4.17 × 10⁻³ and a root mean square error of 2.38 × 10⁻², indicating lower precision compared to the other two methods.

The evaluation revealed that the validity and reliability of IPD reconstruction were lower with the Guyot method compared to the direct and Rogula methods, especially when used on smaller datasets. Errors in metrics using Guyot-derived data were particularly high for cohorts of 20 or 50 subjects, whereas errors seemed to converge with those from the other methods when the Guyot method was used on larger datasets. The authors suggest this discrepancy is caused by the algorithm’s tendency to accumulate errors in survival estimates due to less accurate placement of censorings, which can result in visible deviations at the tail of the curve.

For hazard ratios calculated from data reconstructed using the direct and Rogula methods, 94% and 89% respectively had an absolute error smaller than 0.01, and 59% for both had an error smaller than 0.001. In contrast, only 32% of hazard ratios calculated with IPD estimated using the Guyot method had an absolute error less than 0.01, and only 18% had an error less than 0.001. These thresholds correspond to the limit of detecting differences when comparing with published hazard ratios, which are typically reported with two or three decimal places. The authors considered the overall accuracy of the reconstructed data to be sufficient for all common types of secondary data use.

The study also examined the uncertainties of hazard ratios, measured as the standard error of the log hazard ratio. These uncertainties were higher with Guyot-derived data, with a median absolute error of 1.92 × 10⁻⁴, compared to 4.50 × 10⁻⁶ for the direct method and 7.32 × 10⁻⁶ for the Rogula method. Survival quantiles from reconstructed data, including medians, generally showed very high similarity with their respective references, although errors in landmark survival were slightly higher with data reconstructed using the Guyot method. The findings support the use of the direct and Rogula methods in scenarios where precise metrics are of interest, particularly for smaller datasets.

Several limitations to IPD reconstruction in general and in this specific study remain. The image quality of the underlying figure can significantly influence how easily its data can be reconstructed, and the study used images of good quality without formally assessing the impact of poor-quality images. The degree of censoring is another potential issue; when censoring indicators overlap, the direct and Rogula methods may be more difficult to apply, though they remain usable with some adjustment. In cases of excessive censoring, the Guyot algorithm, which estimates censoring, may be a useful alternative, although it is unclear whether it performs better than the other methods in such settings.

The authors note that neither the Guyot nor the Rogula methods are suitable for reconstructing survival data from Aalen-Johansen plots or Simon-Makuch plots, which are used in multi-state survival or competing risk models and settings with time-varying variables, respectively. Consequently, V-Exploder does not support IPD reconstruction for these specific types of plots. The tool is designed to handle standard Kaplan-Meier plots, which are the most common graphical representation of time-to-event endpoints in clinical trials. The application’s ability to save project files for later use and collaboration may ultimately help improve scientific rigor in IPD reconstruction by allowing for external accuracy assessment and iterative refinement.

V-Exploder is available as a platform-independent web application, with all data related to IPD reconstruction kept locally and not uploaded to a remote server. This ensures the privacy of potentially confidential data while providing a user-friendly interface for researchers. The tool’s integration of the complete workflow, from digitization to final data export, streamlines the process of secondary data analysis. By offering immediate visual feedback and multiple reconstruction methods, V-Exploder facilitates the accurate extraction of individual patient data, supporting a wide range of medical research applications including meta-analyses and cost-effectiveness analyses.

Subject of Research: Development and evaluation of a web-based tool for reconstructing individual patient data from Kaplan-Meier survival plots

Article Title: Visual extraction of individual patient data from published kaplan-meier survival plots using a web-based application offering different reconstruction methods (V-Exploder)

Article References: Acker, F., Ruge, L., Rost, M., Schulte, H., Heinzen, S., Althoff, F. C., Kouidri, K., Stratmann, J., Herrmann, E., Sebastian, M., & Koschade, S. E. (2026). Visual extraction of individual patient data from published kaplan-meier survival plots using a web-based application offering different reconstruction methods (V-Exploder). BMC Medical Research Methodology, 26(1), Article 216. https://doi.org/10.1186/s12874-026-03036-3

Image Credits: AI Generated

DOI: 10.1186/s12874-026-03036-3

Keywords: Kaplan-Meier plots, individual patient data, survival analysis, V-Exploder, data reconstruction, medical research, web application, meta-analysis, cost-effectiveness, statistical methods, Visual, extraction