# TrialCurve

> TrialCurve turns an image of a published Kaplan–Meier survival curve into digitised curves, the numbers-at-risk table, reconstructed individual patient data (pseudo-IPD, Guyot et al. 2012 algorithm), hazard ratios with 95% CIs, medians and RMST — via a web demo or a REST API. It is built for HEOR, market-access and evidence-synthesis teams doing network meta-analysis, MAIC, survival extrapolation and HTA submissions (EU Joint Clinical Assessment, NICE, CDA-AMC, PBAC).

How it works: Claude Opus 5.5 reads the whole figure (every survival panel's axes, arms, curves as step corners, numbers-at-risk table and printed statistics), zooming in wherever the full view is too small. Pixel measurement traces each panel and replaces the model's curve where the two agree. The Guyot algorithm reconstructs patient-level data, and every reconstruction is checked against the hazard ratios, medians and event counts printed on the figure; a figure that disagrees is re-read by a larger configuration. If the vision model is unavailable, a measurement-only engine (on-device OCR and pixel tracing) takes over. Figures are processed in memory and not stored.

Full text of the main pages: https://trialcurve.com/llms-full.txt

## Key pages
- [Home and live demo](https://trialcurve.com/): upload a KM figure or try examples
- [API documentation](https://trialcurve.com/docs): POST /v1/reconstruct, options, response schema, errors
- [Accuracy and methodology](https://trialcurve.com/accuracy): the production pipeline measured on figures from published trials (against their printed statistics) and on held-out simulated figures with known ground truth; known limitations
- [Trial library](https://trialcurve.com/trials): reconstructed Kaplan–Meier curves and pseudo-IPD from published randomised trials (CC BY sources), validated against the printed statistics, with CSV downloads
- [Free academic access](https://trialcurve.com/academic): free API credits for published or registered methods work
- [Contact](https://trialcurve.com/contact): enterprise, invoicing, support; or email hello@trialcurve.com
- [Terms and privacy](https://trialcurve.com/legal)

## Facts
- Output: per-arm curve coordinates, at-risk table, reconstructed IPD (time, event), N, events, median, survival at landmarks, RMST, Cox hazard ratio (Efron ties) with 95% CI, log-rank p; CSV export for R/Stata/SAS.
- Pricing: 3 free figures per month; $15 per figure; audit-grade (validation pack) $45 per figure; Team $490/month including 50 credits; enterprise private deployment available.
- Method reference: Guyot P, Ades AE, Ouwens MJ, Welton NJ. BMC Med Res Methodol 2012;12:9.

## Guides, methods and tools
- [Reconstruct IPD from Kaplan–Meier curves](https://trialcurve.com/reconstruct-ipd)
- [How to extract data from a Kaplan–Meier curve](https://trialcurve.com/guides/extract-data-from-kaplan-meier-curve)
- [Numbers-at-risk tables explained](https://trialcurve.com/guides/number-at-risk-table)
- [The Guyot algorithm (2012), explained](https://trialcurve.com/methods/guyot-algorithm)
- [Hazard ratio from a KM curve: calculator and methods](https://trialcurve.com/tools/hazard-ratio-from-kaplan-meier-curve)
- [RMST calculator](https://trialcurve.com/tools/rmst-calculator)
- [Kaplan–Meier curve generator](https://trialcurve.com/tools/kaplan-meier-curve-generator)
- [How to read a Kaplan–Meier curve](https://trialcurve.com/guides/how-to-read-a-kaplan-meier-curve)
- [Hazard ratio interpretation](https://trialcurve.com/guides/hazard-ratio-interpretation)
- [How to cite TrialCurve](https://trialcurve.com/how-to-cite)
- [Reconstructed IPD in HTA (NICE TSD 14/21, NMA, MAIC, EU JCA)](https://trialcurve.com/hta)
- [KM digitizer and IPD tools compared](https://trialcurve.com/compare/kaplan-meier-digitizer-tools)
- [WebPlotDigitizer alternative for survival curves](https://trialcurve.com/compare/webplotdigitizer-alternative)

## Accuracy (the production pipeline; details at /accuracy)
- Real published figures: 87 whole Kaplan–Meier figures from 40 CC BY randomised trials (154 panels reconstructed). Printed hazard ratios reproduced within 10%: 95% of 141 (within 5%: 82%); printed medians within 10%: 96% of 128. An independent re-read by a second model reproduced all 14 remaining differences, which point to the figures (adjusted or stratified HRs, very small arms).
- Simulated figures with known truth (80, held out): hazard ratio within 5% of truth 91.2% (model reading alone 93.8%, measurement alone 85.5%); within 10% 96.2%; median error in median survival 0.09 months; 100% of at-risk cells read exactly; true HR inside the reconstructed 95% CI in 100% of figures.

## How to cite
TrialCurve (2026). Automated reconstruction of individual patient data from Kaplan–Meier figures. https://trialcurve.com