Letao Li, John Maringwa1, Chandni Valiathan
1Johnson & Johnson Innovative Medicine
Introduction: The treatment landscape in relapsed/refractory multiple myeloma (RRMM) has seen improvement in recent years, with new therapies often receiving accelerated approval based on early response endpoints like overall response rate (ORR), which are available sooner than survival endpoints e.g., progression-free survival (PFS). The correlation between ORR and median PFS (mPFS) quantified the relationship between the endpoints. However, several studies that have reported early endpoints are excluded in such analyses due to immature PFS (mPFS not reached) at the time of publication. For most studies, including those with immature survival data, Kaplan-Meier (KM) data up to a certain time point is usually available, and can be informative. This study investigated if available KM data could reliably be used to project median survival times for immature survival data. Studies that reported both KM curves and mPFS were leveraged. Results from this study may help researchers to supplement reported literature data through imputation of mPFS for immature survival data cases, providing more information for inference, potentially helping improve precision on estimates. This analysis identifies recommendations on the amount of KM curve-based information required to reliably estimate the median survival times for immature survival data to help practitioners decide on when or whether it is advisable to use the imputed information. Methods: A systematic literature review in RRMM that included studies published between 2007 and 2023 for both investigational and approved antimyeloma therapies, involving 140 (186 study arms) studies was used to explore mPFS in relation to ORR. The strength of the relationship was assessed by the proportion of variation in mPFS that can be explained by ORR as quantified by the R-square. Longitudinal models (assuming a Weibull or Exponential-based hazard of progression or death [1] were explored to describe the trajectory of the KM curve for each study arm. Weighting of observations was based on the number of patients at risk when reported, otherwise no censoring was assumed. Subsets of data (e.g., 90%, 80%, 70%, …) were created to reflect the amount of information retained as a fraction of the total expected study duration, after excluding part of data to mimic “immature” survival data scenarios. The reported KM data represented 100% information available. Median PFS estimates projected from the parametric models across data subsets were compared with the “true” reported mPFS. Plots of reported and model projected mPFS, median prediction error (MPE) and root mean square error (RMSE) were used to evaluate model adequacy, enabling comparison between the parametric models, as well as the impact of the fraction of information retained. Results: Based on all available KM data, the Weibull and Exponential models demonstrated good fit, achieving R-square estimates of 0.87 and 0.85, respectively, between model projected and reported mPFS. The MPE and RMSE based on all data, as well as subsets of data, were larger for the Exponential model, suggesting the Weibull model provided a better description of the data compared to the Exponential, but the difference became wider with less information available. With 70% of the data retained, R-square estimates were 0.82 and 0.76 for the Weibull and Exponential, respectively. When information was reduced to 50%, correlations estimates were 0.73 and 0.64, for the Weibull and Exponential, respectively. Conclusion: These findings highlight the possibility of early efficacy indicator ORR to serve as a potential surrogate endpoint for PFS as reflected by the relatively strong correlation. Further, the results suggest that parametric models e.g., Weibull or Exponential can reasonably project median survival times in case of immature survival data. Relatively high correlation was achieved with about 70% of the KM curve data, suggesting that when at least 70% of information is available, one may be able to reasonably project unreported median PFS values when faced with immature survival data. This can help to enrich the analysis dataset compared to excluding those studies for which the median PFS had not been reached at the time of reporting, potentially avoiding bias.
1. Collett, D. (2014). Modelling survival data in medical research (3rd ed.). Chapman & Hall/CRC.
Reference: PAGE 33 (2025) Abstr 11563 [www.page-meeting.org/?abstract=11563]
Poster: Drug/Disease Modelling - Oncology