Kaplan-Meier reconstruction check

Digitized versus reconstructed survival curve

MAIC

STC

ML-NMR

ML-UMR

NMA

A published Kaplan-Meier curve overlaid with the curve rebuilt from reconstructed pseudo-IPD, to validate the reconstruction.

MAIC STC ML-NMR ML-UMR NMAEstablished

Overall survival with levamisole plus 5-FU in a colon cancer trial: the original Kaplan-Meier curve (solid) and the curve from pseudo-IPD reconstructed with the Guyot algorithm from the curve’s coordinates and yearly numbers at risk (dashed). Data: survival::colon.

Family

Survival and time-to-event

Purpose

Validate pseudo-IPD before using it in MAIC, STC, ML-NMR, ML-UMR, or survival NMA.

Inputs

Digitized curve coordinates and published numbers at risk.

Software

R IPDfromKM (shown), survHE::digitise(); Stata ipdfc

What it shows

Population-adjusted and survival syntheses often need patient-level survival data for a comparator that has only been published as a Kaplan-Meier curve. The Guyot algorithm reconstructs pseudo-IPD from digitized curve coordinates and the numbers at risk. Plotting the reconstructed Kaplan-Meier curve over the original, and comparing medians and hazard ratios where published, is the standard check that the reconstruction is faithful.

How to read it

Solid curve: the original published curve (here, the true Kaplan-Meier from the trial data).
Dashed curve: Kaplan-Meier estimate from the reconstructed pseudo-IPD.
Agreement: the curves should overlap closely at every time.

Interpretation

With exact coordinates and a yearly risk table, the reconstruction is essentially perfect: the two curves overlap throughout, and 5-year survival (63%) is reproduced. Real digitization introduces error, particularly where the published curve is steep or the risk table sparse, so the check matters more in practice than here.

Pitfalls

Reconstruction quality depends on the resolution of the digitized figure and on the numbers at risk; without a risk table, censoring must be assumed.
Pseudo-IPD contain no covariates; they cannot support adjustment on their own.
Report the reconstructed medians and hazard ratios against the published ones.

Code

library(survival)
library(IPDfromKM)
library(ggplot2)

# Start from a known Kaplan-Meier curve (colon cancer trial, levamisole + 5-FU)
data(colon, package = "survival")
arm <- subset(colon, etype == 2 & rx == "Lev+5FU")
km <- survfit(Surv(time / 365.25, status) ~ 1, data = arm)

# "Digitize" it: coordinates at the steps, plus numbers at risk every year,
# as would be read from a published figure
coords <- data.frame(time = km$time, surv = km$surv)
trisk <- 0:8
nrisk <- summary(km, times = trisk, extend = TRUE)$n.risk

# Guyot algorithm: reconstruct pseudo-IPD from the curve and risk table
prep <- preprocess(dat = coords, trisk = trisk, nrisk = nrisk, maxy = 1)
ipd <- getIPD(prep = prep, armID = 1)$IPD
rec <- survfit(Surv(time, status) ~ 1, data = ipd)

curves <- rbind(
  data.frame(time = c(0, km$time), surv = c(1, km$surv), source = "Original Kaplan-Meier"),
  data.frame(time = c(0, rec$time), surv = c(1, rec$surv), source = "Reconstructed from digitized curve")
)
ggplot(curves, aes(time, surv, colour = source, linetype = source)) +
  geom_step(linewidth = 0.9) +
  scale_colour_manual(values = c("#1b1f24", "#b5452b"), name = NULL) +
  scale_linetype_manual(values = c("solid", "dashed"), name = NULL) +
  scale_y_continuous(labels = scales::percent, limits = c(0, 1)) +
  labs(x = "Years", y = "Overall survival",
       title = "Checking a Kaplan-Meier reconstruction",
       subtitle = "Original curve against pseudo-IPD reconstructed with the Guyot algorithm") +
  guides(colour = guide_legend(ncol = 1))

References

Guyot P, Ades AE, Ouwens MJNM, Welton NJ. Enhanced secondary analysis of survival data: reconstructing the data from published Kaplan-Meier survival curves. BMC Med Res Methodol. 2012;12:9. doi:10.1186/1471-2288-12-9
Liu N, Zhou Y, Lee JJ. IPDfromKM: reconstruct individual patient data from published Kaplan-Meier survival curves. BMC Med Res Methodol. 2021;21:111. doi:10.1186/s12874-021-01308-8