Reconstruct individual patient data from published Kaplan–Meier curves

Updated 2026-09-25 · TrialCurve methods team

TrialCurve does all of it from a single image. Claude Opus 5.5 reads the whole figure: every panel's axes, arm names, curves, numbers-at-risk table and printed statistics, zooming in wherever the full view is too small. Where pixel measurement of a panel agrees with that reading, the measured curve is used. Each arm is then reconstructed, and the result is checked against the statistics printed on the figure.

What you get back

• Digitised curves for every arm, with censor-mark positions.
• The numbers-at-risk table, read cell by cell by the vision model, which zooms into the table wherever the digits are small.
• Pseudo-IPD as one row per reconstructed patient (arm, time, event), ready for survival, flexsurv, multinma or Stata's stcox.
• Summary statistics: N, events, median survival, survival at landmarks, RMST, Cox hazard ratio with 95% CI and log-rank p.
• A validation pack: reconstruction versus the values printed on the figure (HR, medians, events), agreement between two independent extractions, a QC overlay and provenance (image hash, engine version).

How the reconstruction works

The Guyot algorithm treats the digitised curve as a series of step points and the at-risk table as fixed totals at the start of each interval. Within each interval it estimates how many patients were censored so that the implied number at risk at the next table time matches the published count. It assumes censoring is spread evenly within the interval. It then assigns events at each step so that the Kaplan–Meier estimate matches the curve's drop. The final interval uses the average censoring rate, or the total number of events if the paper reports it. See our explainer of the Guyot algorithm for the details.

How accurate is it?

When the at-risk table is available, reconstruction accuracy is limited mainly by how precisely the curve is digitised. On 80 held-out synthetic journal-style figures with known ground truth, the production engine put 93.6% of hazard ratios within 10% of the true value (88.5% within 5%), with a median error in median survival of 0.10 months. For comparison, a validation of manual digitisation plus Guyot found a mean HR error of 0.0094 (Saluja et al. 2019). Full methodology and failure modes are on the accuracy page.

When reconstruction is less reliable

• No numbers-at-risk table. Pass N per arm, and total events if reported. Expect wider error in the tail.
• Crossing curves of the same colour, heavy compression or tiny screenshots. Check the overlay and the quality grade.
• Heavy censoring between table times, where the even-censoring assumption is strained.

Using pseudo-IPD downstream

Common uses include survival extrapolation for cost-effectiveness models (NICE DSU TSD 14, TSD 21), checking proportional hazards, network meta-analysis of time-to-event outcomes, and unanchored or anchored MAIC (TSD 18). The HTA guide covers how assessors view reconstructed data.

FAQ