Figure format and the accuracy of AI-assisted reconstruction of Kaplan-Meier survival data
- 1The University of Arizona
Abstract
Reconstructing participant-level data from published Kaplan-Meier curves supports meta-analysis and other secondary survival analyses when original trial records are unavailable. Researchers may use original PDF figures or screenshots, which can look alike but retain different information. We assessed how converting vector figures to images affects reconstruction by a language model that can inspect files and execute code. We generated 64 independent two-arm datasets across four scenarios combining 100 or 1,000 participants per arm with proportional or crossing hazards. Each figure was supplied as a native vector PDF and a matched 300-dpi PNG to Opus 5.5 through Claude Code under the same instructions. Outcomes compared reconstructed records with the original observed records: time-averaged absolute curve error through 24 months, error in the between-group restricted mean survival time difference, and error in the Cox log-hazard ratio. Native PDFs had lower curve error in all 64 completed pairs, including replacements for incomplete runs. Mean errors were 0.000029 and 0.1047 percentage points for PDF and PNG, respectively; the paired difference was -0.1047 (95% confidence interval, -0.1194 to -0.0909). PNG errors in the restricted mean survival time difference and log-hazard ratio were small: mean absolute errors were 0.0084 months and 0.0027. Lower resolution and compression increased curve errors; embedding unchanged pixels in a PDF did not restore vector accuracy. Some PNG runs exceeded the instructed revision limit, so results describe the implemented workflow rather than performance under a strictly enforced limit. Preserving vector information improved reconstruction fidelity in this workflow, but the observed differences do not establish clinically consequential changes in treatment estimates. Validation against reported survival summaries and event counts remains necessary. Generalization to published clinical figures and other model workflows requires further evaluation.
Related
- Navigation-Assisted and Robotic Total Hip, Knee, and Unicompartmental Arthroplasty: Reconciling Positioning Accuracy with Biomaterial Performance and Long-Term Implant Survival.
- Traumatic Left Posterior Sternoclavicular Joint Dislocation with Figure-of-eight Semitendinosus Ligament Reconstruction
- AI-Assisted Triage for Diabetic Retinopathy Integrated into Routine Diabetes Care: A Real-World Diagnostic Accuracy Study of the DAIRET® Algorithm
- AI-assisted 3D thoracic body composition analysis identifies candidate imaging markers associated with overall survival in older patients with lung cancer receiving radiotherapy
- Visor Flap-Assisted Segmental Mandibulectomy and Immediate Free Fibula Reconstruction for Extensive Plexiform Ameloblastoma of the Anterior Mandible: A Case Report
BibTeX
@misc{li2026figure,
title = {{Figure format and the accuracy of AI-assisted reconstruction of Kaplan-Meier survival data}},
author = {S. Li and X. Xing and B. Te and H. Han and H. Ji and Y. Yu and L. Siegel and L. Lin},
howpublished = {medRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.03.26367147},
url = {https://www.medrxiv.org/content/10.64898/2026.11.03.26367147v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Figure format and the accuracy of AI-assisted reconstruction of Kaplan-Meier survival data
AU - S. Li
AU - X. Xing
AU - B. Te
AU - H. Han
AU - H. Ji
AU - Y. Yu
AU - L. Siegel
AU - L. Lin
PY - 2026
DA - 2026/11/03
PB - medRxiv
DO - 10.64898/2026.11.03.26367147
UR - https://www.medrxiv.org/content/10.64898/2026.11.03.26367147v1
AB - Reconstructing participant-level data from published Kaplan-Meier curves supports meta-analysis and other secondary survival analyses when original trial records are unavailable. Researchers may use original PDF figures or screenshots, which can look alike but retain different information. We assessed how converting vector figures to images affects reconstruction by a language model that can inspect files and execute code. We generated 64 independent two-arm datasets across four scenarios combining 100 or 1,000 participants per arm with proportional or crossing hazards. Each figure was supplied as a native vector PDF and a matched 300-dpi PNG to Opus 5.5 through Claude Code under the same instructions. Outcomes compared reconstructed records with the original observed records: time-averaged absolute curve error through 24 months, error in the between-group restricted mean survival time difference, and error in the Cox log-hazard ratio. Native PDFs had lower curve error in all 64 completed pairs, including replacements for incomplete runs. Mean errors were 0.000029 and 0.1047 percentage points for PDF and PNG, respectively; the paired difference was -0.1047 (95% confidence interval, -0.1194 to -0.0909). PNG errors in the restricted mean survival time difference and log-hazard ratio were small: mean absolute errors were 0.0084 months and 0.0027. Lower resolution and compression increased curve errors; embedding unchanged pixels in a PDF did not restore vector accuracy. Some PNG runs exceeded the instructed revision limit, so results describe the implemented workflow rather than performance under a strictly enforced limit. Preserving vector information improved reconstruction fidelity in this workflow, but the observed differences do not establish clinically consequential changes in treatment estimates. Validation against reported survival summaries and event counts remains necessary. Generalization to published clinical figures and other model workflows requires further evaluation.
ER -