Figure format and the accuracy of AI-assisted reconstruction of Kaplan-Meier survival data

Native vector PDFs enabled more accurate AI-assisted reconstruction of Kaplan-Meier survival curves than matched 300-dpi PNGs, with mean time-averaged curve error 0.000029 vs 0.1047 percentage points across 64 datasets.

• Why it matters: When original trial records are missing, reconstructing participant-level survival data from published curves is needed for secondary analyses, but figure conversion can silently change extractable information.
• What they did: The study generated 64 two-arm datasets (100 or 1,000 per arm; proportional or crossing hazards) and provided each as a native vector PDF and matched 300-dpi PNG to Opus 5.5 via Claude Code under the same instructions, then compared reconstructions to observed records.
• The result: PDFs reduced curve error in all completed pairs, with a paired difference of -0.1047 (95% CI -0.1194 to -0.0909); PNG-related errors also remained small for restricted mean survival time difference (0.0084 months) and Cox log-hazard ratio (0.0027), implying vector preservation improves fidelity in this workflow.

AI summary of the abstract · may contain errors · not medical advice · read the paper.