Kseniia Danichkina1, Kirill Zhudenkov2,3, Kirill Peskov2,3,4, Ekaterina Khrameeva1
1Skolkovo Institute of Science and Technology, 2M&S Decisions FZ-LLC, 3I.M. Sechenov First Moscow State Medical University, 4Marchuk Institute of Numerical Mathematics RAS
Introduction: Survival analysis is a gold standard of treatment efficacy assessment in oncology clinical trials. Individual patient data (IPD) are integrated to construct Kaplan-Meier (KM) curves representing published data for patient survival [1]. The basic assumptions used in survival analysis are the proportionality of hazards (PH) (change in probability of event due to change in covariate does not depend on time) and non-informativeness of censoring (dropouts occur completely at random or at least with similar intensity across arms compared) [2]. However, this is not always the case in various clinical trials [3-5]. Here, we investigate and compare available methods of published KM curves digitization and IPD retrieval [6-9]. Objectives: The objectives of the presented work included: a development of a seamless workflow for IPD data digitization and analysis, e.g. identification and comparison of efficient algorithms for the published KM curves in clinical trials; application the diagnostic tests to check the proportionality of hazards to verify the consistency of rank tests and hazard ratio (HR) assessment; performing imbalanced censoring tests and sensitivity analysis to identify possible bias in survival comparison outcomes. Methods: Representative survival data were obtained from a set of published clinical trials in advanced NSCLC [10-12]. Pairs of KM curves from different patient subgroups, representing comparison of investigated drugs and control with various sample sizes, were digitized using the PlotDigitizer [13]. The Liu IPD reconstruction algorithm was used with and without information on the risk table [7]. The Rogula IPD method was implemented in three scenarios [6]: with exact censoring times, with uniformly distributed censoring times and without censoring times. Reconstruction adequacy implied the overlay of original and reconstructed KM curves. Median survival times, published HRs and their uncertainty were also compared. The PH assumption (as well as the reliability of the log-rank test and HR estimates) was verified using the Schoenfeld, Martingale and Deviance residuals and compared with published results. Imbalanced censoring tests for reconstructed IPD and sensitivity analysis were performed using the reverse KM method described in Gilboa et al. [5]. All analyses were conducted in R 4.3.3 [14]. Results: In 3 selected clinical studies, 13 pairs of KM curves were taken for IPD reconstruction and diagnostics, representing different sample sizes (N=20-425) and curve shapes [10-12]. Survival curves generated from the reconstructed IPD agreed well with published KM curves for the Liu and Rogula methods. The outcomes of rank comparison tests and estimated HRs (0.86 (0.71-1.06), 0.87 (0.71-1.06), 0.88 (0.72-1.07); 0.77 (0.65-0.92), 0.78 (0.66-0.93), 0.77 (0.65-0.92); 0.73 (0.62-0.87), 0.73 (0.61-0.87), 0.67 (0.57-0.78) for published, Liu and Rogula method results within 95%CI for the overall survival estimates in 3 clinical studies [10-12], respectively). However, hazard proportionality was violated in 3 pairs of the KM curves, resulting in a possible bias in the rank comparison test results as well as HR (both published and estimated from reconstructed IPD). Censoring diagnostics and sensitivity analysis by means of reverse KM method revealed 1 case of imbalanced censoring in the reconstructed IPD for both methods and 4 additional imbalanced censoring signals for the IPD reconstructed using the less robust Liu method (the method does not require precise information on censoring events). These results indicate that, in certain cases, patient dropout followed apparently different patterns. This may be due to details of the safety profile or other differences associated with the compared therapies. Conclusions:? Here we presented a workflow for the IPD reconstruction along with necessary inspections for the data prior to the qualification of survival models. Both tested methods for IPD reconstruction showed high performance at a qualitative level for various sample sizes. The Rogula method was more robust, though it required more information on censoring of subjects. In certain cases, the data showed a violation of hazard proportionality. Thus, in further analysis, the advanced survival models should be developed [15]. Finally, the analysis of censoring revealed additional differences in treatment effects that would require further application of competing risk models and other advanced approaches.
1. Fojo T and Simon R (2021). Inappropriate censoring in Kaplan-Meier analyses. Lancet Oncol, 22(10), 1358 – 1360. 2. Emmerson J and Brown J (2021). Understanding Survival Analysis in Clinical Trials. Clinical Oncology, 33(1), 12 – 14. 3. Rosen K et al. (2020). Censored patients in Kaplan–Meier plots of cancer drugs: An empirical analysis of data sharing. Eur J Cancer, 141, 152 – 161. 4. Templeton A et al. (2020). Informative censoring – a neglected cause of bias in oncology trials. Nat Rev Clin Oncol, 17(6), 327 – 328. 5. Gilboa S et al. (2021). Informative censoring of surrogate end-point data in phase 3 oncology trials. Eur J Cancer, 153, 190-202. 6. Rogula B et al. (2022). A Method for Reconstructing Individual Patient Data From Kaplan-Meier Survival Curves That Incorporate Marked Censoring Times. MDM Policy Pract, 7(1), 23814683221077643. 7. Liu N et al. (2021). IPDfromKM: reconstruct individual patient data from published Kaplan-Meier survival curves. BMC Med Res Methodol, 21(1), 111. 8. Guyot P et al. (2012). Enhanced secondary analysis of survival data: reconstructing the data from published Kaplan-Meier survival curves. BMC Med Res Methodol, 12, 9 9. Wei Y and Royston P (2017). Reconstructing time-to-event data from published Kaplan-Meier curves. Stata J., 17(4), 786 – 802. 10. Nishio M et al. (2021). Atezolizumab Plus Chemotherapy for First-Line Treatment of Nonsquamous NSCLC: Results From the Randomized Phase 3 IMpower132 Trial. J Thorac Oncol, 16(4), 653–664. 11. Johnson M et al. (2023). Durvalumab With or Without Tremelimumab in Combination With Chemotherapy as First-Line Therapy for Metastatic Non–Small-Cell Lung Cancer: The Phase III POSEIDON Study. J Clin Oncol, 41(6), 1213–1227. 12. Rittmeyer A et al. (2017). Atezolizumab versus docetaxel in patients with previously treated non-small-cell lung cancer (OAK): a phase 3, open-label, multicentre randomised controlled trial. Lancet, 389(10066), 255–265. 13. PlotDigitizer website: https://plotdigitizer.sourceforge.net/ 14. R Core Team (2024). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/ 15. Bardo M et al. (2024). Methods for non-proportional hazards in clinical trials: A systematic review. Statistical Methods in Medical Research, 33(6), 1069-1092.
Reference: PAGE 33 (2025) Abstr 11390 [www.page-meeting.org/?abstract=11390]
Poster: Methodology - Other topics