A Web-Based Application To Automate the Development of State-Transition Models in Oncology Using Published Survival Curves and Bayesian Calibration
Author(s)
David U. Garibay Treviño, BA, MSc1, Alexandra Moskalewicz, M.Sc., B.Sc.2, Muhammad Enrizky Brillian, B.Sc.3, Qiyue Zhang, B.Sc.4, Matthew Kauffmann, Ph.D.5, Eline Krijkamp, Ph.D.6, Fernando Alarid-Escudero, Ph.D.7, Hawre Jalal, Ph.D.8, Eva Enns, Ph.D.5, Petros Pechlivanoglou, Ph.D.2;
1University of Ottawa, School of Epidemiology and Public Health, Ottawa, ON, Canada, 2The Hospital for Sick Children Research Institute, Toronto, ON, Canada, 3University of Toronto, Toronto, ON, Canada, 4The University of Toronto, Toronto, ON, Canada, 5University of Minnesota, Minneapolis, MN, USA, 6Erasmus University, Rotterdam, Netherlands, 7Stanford University, Stanford, CA, USA, 8University of Ottawa, Ottawa, ON, Canada
1University of Ottawa, School of Epidemiology and Public Health, Ottawa, ON, Canada, 2The Hospital for Sick Children Research Institute, Toronto, ON, Canada, 3University of Toronto, Toronto, ON, Canada, 4The University of Toronto, Toronto, ON, Canada, 5University of Minnesota, Minneapolis, MN, USA, 6Erasmus University, Rotterdam, Netherlands, 7Stanford University, Stanford, CA, USA, 8University of Ottawa, Ottawa, ON, Canada
Presentation Documents
OBJECTIVES: Modelers can encounter challenges building state-transition models (STM) when individual-patient data (IPD) is not available. We developed a web-based application to automate parameterization of three-state STMs in oncology using published survival curves and bayesian calibration.
METHODS: Within the web-based application (https://pechlilab.shinyapps.io/STM_calibration/), users first upload images of overall survival (OS) and progression-free survival (PFS) curves. Utilizing a combination of R packages, the survival curves are digitized, IPD is reconstructed, and Kaplan-Meier (KM) curves are regenerated to define calibration targets. Assuming a three-state framework (progression-free, progressed, death), Bayesian calibration is performed to obtain a posterior distribution of parameters (n=1,000) for each transition.
We illustrate this approach using IPD from a clinical trial in colon cancer (n=315, observation arm, 9-year max follow-up). To initialize the calibration, the first 3 years of the “true” IPD were used to generate images of KM curves for OS/PFS. Resulting calibrated parameter sets were used to fit a semi-markov STM with a 9-year time horizon. Two alternative modeling approaches were evaluated against the true IPD: a partitioned survival model (PSM) informed by the 3-year digitized OS/PFS curves, and a multi-state model (MSM) fit with 3-year true IPD. Performance was estimated using root mean squared errors (RMSE) by comparing model-generated OS/PFS against the true IPD.
RESULTS: From the web application, users can obtain parameter correlation plots, validation plots, and calibrated parameter sets with likelihood values. After applying the colon dataset, OS estimates from the true IPD most closely matched the calibrated STM (RMSE: 0.067), followed by the PSM (0.089), and MSM (0.111). RMSE values for PFS were 0.070 (calibrated STM), 0.081 (PSM), and 0.100 (MSM).
CONCLUSIONS: Bayesian calibration can be applied to accurately inform STMs when only published survival curves are available.
METHODS: Within the web-based application (https://pechlilab.shinyapps.io/STM_calibration/), users first upload images of overall survival (OS) and progression-free survival (PFS) curves. Utilizing a combination of R packages, the survival curves are digitized, IPD is reconstructed, and Kaplan-Meier (KM) curves are regenerated to define calibration targets. Assuming a three-state framework (progression-free, progressed, death), Bayesian calibration is performed to obtain a posterior distribution of parameters (n=1,000) for each transition.
We illustrate this approach using IPD from a clinical trial in colon cancer (n=315, observation arm, 9-year max follow-up). To initialize the calibration, the first 3 years of the “true” IPD were used to generate images of KM curves for OS/PFS. Resulting calibrated parameter sets were used to fit a semi-markov STM with a 9-year time horizon. Two alternative modeling approaches were evaluated against the true IPD: a partitioned survival model (PSM) informed by the 3-year digitized OS/PFS curves, and a multi-state model (MSM) fit with 3-year true IPD. Performance was estimated using root mean squared errors (RMSE) by comparing model-generated OS/PFS against the true IPD.
RESULTS: From the web application, users can obtain parameter correlation plots, validation plots, and calibrated parameter sets with likelihood values. After applying the colon dataset, OS estimates from the true IPD most closely matched the calibrated STM (RMSE: 0.067), followed by the PSM (0.089), and MSM (0.111). RMSE values for PFS were 0.070 (calibrated STM), 0.081 (PSM), and 0.100 (MSM).
CONCLUSIONS: Bayesian calibration can be applied to accurately inform STMs when only published survival curves are available.
Conference/Value in Health Info
2025-05, ISPOR 2025, Montréal, Quebec, CA
Value in Health, Volume 28, Issue S1
Code
MSR134
Topic
Methodological & Statistical Research
Disease
No Additional Disease & Conditions/Specialized Treatment Areas, SDC: Oncology