VISION-GUIDED RECONSTRUCTION OF INDIVIDUAL PATIENT DATA FROM KAPLAN-MEIER CURVES: HELD-OUT VALIDATION OF AN ASSISTED DIGITIZER WITH CALIBRATED CONFIDENCE
Author(s)
Máté Szilcz, PhD1, Sergio Enmanuel Flores, MASc, MSPH, MD, PhD2.
1Founder & CEO @Viti Science, Karolinska Institutet, Stockholm, Sweden, 2Uppsala University, Uppsala, Sweden.
1Founder & CEO @Viti Science, Karolinska Institutet, Stockholm, Sweden, 2Uppsala University, Uppsala, Sweden.
OBJECTIVES: Reconstructing individual patient data (IPD) from published Kaplan-Meier (KM) curves underlies survival extrapolation, cost-effectiveness analysis, and indirect treatment comparison, yet it is usually done by hand. Manual digitizers require the analyst to place every point and report no measure of reliability, so one figure can give different numbers in different hands. We built a vision-guided digitizer that automates the extraction and reports its own confidence.
METHODS: A multimodal vision-language model proposed axis positions and per-curve anchors from the figure. Computer vision and optical character recognition located the axes and tick labels and traced each curve, separating arms by colour. The pipeline computed a per-axis calibration confidence and flagged unreliable points for review. IPD were reconstructed using the Guyot et al. (2012) method. Synthetic figures with known coordinates were scored by each curve's maximum deviation in months and survival probability. Held-out real figures were checked against printed source statistics where available, otherwise by overlay audit against the source. We tested whether high-confidence outputs agreed with the source and whether errors were flagged.
RESULTS: On six synthetic figures with known truth, every curve fell within 1.5 months and 0.03 survival probability. Arm detection was correct in all 18 development cases. Across 24 held-out overlays audited against the source, no high-confidence result was wrong, and harder figures were flagged for review. On a held-out figure that printed its own statistics, reconstructed median recurrence-free survival was 131.7 versus 133.0 months, and a reconstructed Cox hazard ratio was consistent with the published estimate, which lay within the reconstructed confidence interval. A further 91 figures ran without unflagged errors.
CONCLUSIONS: The tool reproduced the published survival statistics to within about one month on medians, while adding calibrated confidence and a reviewable record that hand digitizers lack. Heavily overlapping and multi-panel figures still need manual correction.
METHODS: A multimodal vision-language model proposed axis positions and per-curve anchors from the figure. Computer vision and optical character recognition located the axes and tick labels and traced each curve, separating arms by colour. The pipeline computed a per-axis calibration confidence and flagged unreliable points for review. IPD were reconstructed using the Guyot et al. (2012) method. Synthetic figures with known coordinates were scored by each curve's maximum deviation in months and survival probability. Held-out real figures were checked against printed source statistics where available, otherwise by overlay audit against the source. We tested whether high-confidence outputs agreed with the source and whether errors were flagged.
RESULTS: On six synthetic figures with known truth, every curve fell within 1.5 months and 0.03 survival probability. Arm detection was correct in all 18 development cases. Across 24 held-out overlays audited against the source, no high-confidence result was wrong, and harder figures were flagged for review. On a held-out figure that printed its own statistics, reconstructed median recurrence-free survival was 131.7 versus 133.0 months, and a reconstructed Cox hazard ratio was consistent with the published estimate, which lay within the reconstructed confidence interval. A further 91 figures ran without unflagged errors.
CONCLUSIONS: The tool reproduced the published survival statistics to within about one month on medians, while adding calibrated confidence and a reviewable record that hand digitizers lack. Heavily overlapping and multi-panel figures still need manual correction.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR171
Topic
Methodological & Statistical Research
Disease
No Additional Disease & Conditions/Specialized Treatment Areas