(SLA-067) A fully automated AI-enabled pipeline for extracting Kaplan-Meier survival data from scientific plots

Background: The ability to extract quantitative data from scientific plots at scale unlocks new possibilities for evidence synthesis and real-world evidence research, particularly when original datasets are unavailable. In clinical research, Kaplan-Meier survival curves are widely used to present time-to-event data. However, existing digitisation methods often require substantial manual effort, limiting their scalability and reproducibility.

Objectives: To develop and validate a computer vision pipeline that automates plot digitisation, enabling large-scale extraction of survival data with minimal human intervention.

Methods: We developed and validated an automated computer vision pipeline using convolutional neural networks, vision-enabled large language models, and edge detection methods. The validation dataset comprised published Kaplan-Meier plots from peer-reviewed scientific literature published between 2020-2024.
Primary outcomes were accuracy metrics including root mean square error (RMSE), mean absolute error, and maximum absolute error between automated and manual digitisation. Secondary outcomes included comparison of survival estimates and Bland-Altman plot analysis.
Performance was evaluated against pre-specified criteria from literature: RMSE ≤0.05, mean absolute error ≤0.02, maximum absolute error ≤0.05, and median survival time within 10% of published estimates. Metrics were adjusted for comparability across different plot scales and formats.

Results: 29 curves across 9 real-world plots were evaluated and digitised both manually and using our automated pipeline. Between manually and automatically digitised coordinates, average RMSE was 0.13 (SD [0.06]) average mean absolute error was 0.11 (SD [0.06]), and average maximum absolute error was 0.28 (SD [0.12]). The average median survival time from manually digitised curves diverged -0.19% from the published median survival time, compared to 1.86% when using our automated pipeline. Though accuracy was lower for automatically extracted coordinates, compared to manually extracted coordinates, our progress represents a significant engineering step towards automation.

Conclusions: Our automated digitisation pipeline demonstrates accuracy approaching that of manual extraction while significantly reducing human effort. The tool's performance across diverse plot styles supports its utility for large-scale data extraction in real-world evidence research, though further validation is needed for heavily annotated or non-standard figures. These results suggest potential for improving reproducibility and efficiency in pharmacoepidemiological research.