How to reconstruct patient data from a Kaplan–Meier plot
- Open the figure: a screenshot of the plot at high zoom (PNG), or the image file. The example shows a finished reading.
- Calibrate the axes: click two labelled ticks on the time axis and two on the survival axis, and type their values.
- Trace each arm: choose the arm, click on its curve, then check the circles against the curve; add or remove points where needed.
- Type the numbers at risk printed under the figure (the times once, then one row per arm). Without them, give the number of patients and, if reported, the events.
- Check and use the result: the dashed reconstructed curve should lie on the published one. Download the data, open it in the Kaplan–Meier tool, or enter the HR and its 95% CI in the forest plot.
Tips for an accurate reading
- Use the largest, sharpest image you can get; zoom into the PDF before taking the screenshot.
- Trace the steps up to the last time in the numbers-at-risk table; censoring ticks and dashed lines are handled.
- Where two curves overlap, add the hidden steps by hand, or remove points that jumped to the other curve.
- Continuous outcomes reported as medians and ranges can be converted with the median to mean and SD calculator.
Frequently asked questions
What does reconstructing IPD from a Kaplan–Meier curve mean?
Papers rarely share their patient-level data, but the Kaplan–Meier figure shows when events happened, and the table under it shows how many patients were still followed. The algorithm of Guyot et al. (2012) combines both to rebuild a data set of event and censoring times whose Kaplan–Meier curve matches the published one. From it you can calculate medians, hazard ratios and restricted mean survival time, or pool trials in a meta-analysis.
How accurate is it?
With the numbers at risk, reconstructed hazard ratios and medians are usually within a few percent of the originals (Guyot et al. 2012; Liu et al. 2021). Without them, the censoring has to be guessed and the error is larger. The tool shows the reconstructed curve over the image, the error against your points (RMSE and maximum), and the reconstructed numbers at risk next to the reported ones, so you can see how well it worked.
Are the results the same as in R?
Yes. The reconstruction is the IPDfromKM package’s implementation of the Guyot algorithm (preprocess and getIPD), and every reconstructed patient matches IPDfromKM 0.1.10 in our reference cases. The R code button gives a script that repeats the reconstruction in R with your digitized points.
Can I use points from WebPlotDigitizer?
Yes. Paste the exported coordinates (CSV or tab-separated time and survival) into the points box of each arm. You can also trace in this tool: calibrate the two axes with four clicks, then click on a curve and it is followed by its colour, including dashed curves and curves with censoring ticks. Where two curves of the same colour meet or cross, check the circles: the tracer stops rather than guess, and you can add the missing points by hand.
How do I use the result in a meta-analysis?
With two or more arms the tool gives the log-rank test and the hazard ratio with its 95% CI, which go into the forest plot tool as estimate and CI, plus the log hazard ratio and its standard error for generic inverse-variance pooling in RevMan or R. “Open in the Kaplan–Meier tool” sends the reconstructed data to that tool for a publication-ready figure. Report that the data were reconstructed, and cite Guyot et al. and IPDfromKM: the methods text does this for you.
Is the image uploaded?
No. The image is read in your browser and the reconstruction runs there too; nothing is sent to a server, which matters for unpublished trial reports and conference slides.