Abstract
Opperman et al. recently published an article: “Predicting Median Survival Times from Early Follow-Up Data in Immunotherapy Trials Using Exponential-Decay Modeling.”1 Related efforts to predict future survival outcomes from limited follow-up data have been reported in other clinical and methodological settings.2-5 In their study, the authors applied the exponential-decay model6 to predict the median overall survival (OS) and the median progression-free survival (PFS) using early portions of Kaplan-Meier (KM) curves, specifically information available by the time the estimated survival probability reached 75%, across 114 phase 3 immunotherapy trials. They reported that the method performed reasonably well for predicting median OS in immunotherapy arms but tended to overestimate median OS in non-immunotherapy arms. Median PFS was overestimated in both immunotherapy and non-immunotherapy arms. When follow-up is limited, the median OS or PFS is often not reached. A simple method for projecting future median survival from early observed survival information could therefore be useful for interpreting emerging trial results, informing clinical decision making, and planning subsequent studies. The proposed approach is attractive because it is transparent, rapid, and readily implementable. However, its simplicity also raises several important methodological and interpretive issues. In medical research, particularly in randomized clinical trials comparing an investigational therapy with a standard treatment, hypothesis testing has become the dominant approach. P-values and statistical significance have become the common currency for evaluating whether a new therapy is effective. However, a P-value <.05 indicates only that the observed data would be relatively unlikely under the null hypothesis; it does not, by itself, establish efficacy or the magnitude or clinical importance of the treatment effect. For better clinical interpretation, estimation is often more informative than hypothesis testing alone. For example, rather than asking only whether OS differs statistically between treatments, clinicians and patients also want to know the magnitude of benefit: How many additional months of median survival time are associated with the investigational treatment? How precise is that estimate? Is the benefit likely to be clinically meaningful? The study by Opperman et al. incorporated both estimation and prediction of subsequent median survival times. Nevertheless, prediction differs fundamentally from estimation based on mature observed data. A predicted median survival time should be accompanied not only by a point estimate but also by an appropriate measure of prediction uncertainty. For time-to-event outcomes such as OS and PFS, survival data are commonly displayed using KM curves and summarized using median survival times. The median is clinically interpretable and often convenient for communication. Under an exponential survival model, the median survival time is directly related to the constant hazard rate. However, the exponential assumption may not adequately represent survival patterns in many oncology trials. Survival data may exhibit nonconstant hazards, delayed treatment effects, waning treatment effects, long-term survival plateaus, or cure fractions. In such settings, a single median survival estimate may fail to capture clinically important aspects of the treatment effect. Median survival is especially inadequate when survival curves cross or when a cure-model framework is more appropriate. In these situations, the restricted mean survival time (RMST) may provide more meaningful and robust summaries of treatment benefit.7,8 The authors state that their method is intentionally simple, interpretable, and rapid and that it may provide a practical way of estimating outcomes before conventional medians are reached. This simplicity is a strength. However, the resulting predicted median appears to be presented primarily as a point estimate without an accompanying measure of uncertainty. A prediction derived from a small trial with few observed events should not be interpreted with the same confidence as a prediction derived from a large trial with more stable early survival estimates. Similarly, uncertainty in digitizing the KM curve, estimating the time corresponding to 75% survival, and specifying the extrapolation model should be reflected in the uncertainty of the predicted median. Several published methods allow reconstruction of approximate individual participant data from published KM curves.9,10 After reconstructed individual-level time-to-event data are available, parametric or nonparametric bootstrap methods can be used to quantify uncertainty in predicted medians. Alternatively, likelihood-based or Bayesian survival models can propagate uncertainty from the observed early survival data into the predicted median. The differences in predictive performance reported by Opperman et al. are potentially informative. The authors suggest that differences in prediction accuracy may reflect differences in survival kinetics between immunotherapy and non-immunotherapy treatments. The observed overestimation of median survival could indicate that hazards increase beyond the 75% survival time in a manner less favorable than would be expected under a constant-hazard exponential model. This pattern appeared to be more pronounced for PFS than for OS. The reported performance also varied across tumor types. Interpretation of these differences requires caution because the number of evaluable patients differed across tumor types and because underlying disease biology, treatment class, subsequent therapy, assessment frequency, and follow-up maturity may all influence predictive performance. From a clinical perspective, a prediction difference of approximately 0.2 months for immunotherapy OS may be acceptable. In contrast, prediction differences of 2 months or more for non-immunotherapy OS or for PFS may be clinically consequential, depending on the disease setting and expected treatment benefit. Several additional factors may affect the prediction of median OS or PFS from early follow-up data. First, treatment exposure time was not incorporated into the model. In oncology trials, treatment may be administered for a prespecified number of cycles, until disease progression, until unacceptable toxicity, or for a maximum duration. Treatment duration and dose intensity may influence PFS and OS. Post-progression therapies can affect OS. Models based solely on early survival probabilities cannot account for these treatment-related factors. Second, PFS differs fundamentally from OS because its determination depends on the timing and method of disease assessment. Radiographic evaluation is typically performed at scheduled intervals, for example, after every 2 treatment cycles. Imaging schedules may also differ by disease type, treatment setting, trial protocol, payer constraints, and routine clinical practice. Unless progression is clinically evident, recorded PFS is therefore interval-censored by the timing of radiographic assessments. OS, by contrast, is not subject to the same assessment-schedule dependence. This distinction may partly explain why the prediction of median PFS was less accurate than the prediction of median OS. Third, the digitization of published KM curves introduces measurement error. Image quality, curve overlap, censoring marks, incomplete numbers-at-risk tables, and limited plotting resolution may affect the accuracy of reconstructed survival information. Finally, it remains unclear whether prediction models should be disease-specific, treatment-class-specific, endpoint-specific, or informed by combinations of these factors. Frequentist methods, including P-values, confidence intervals, and null hypothesis significance testing, have historically dominated medical research and regulatory decision-making. Bayesian methods offer a different and intuitive framework by quantifying the parameters of interest through posterior distributions and providing a probability-based approach to predict future outcomes. Bayesian methods are particularly well-suited for prediction from immature survival data. Prior information from completed trials, studies in similar populations, or historical controls may be incorporated with accumulating data from an ongoing trial to obtain posterior distributions for survival parameters. These posterior distributions can then be used to calculate posterior predictive distributions for median OS, median PFS, RMST, survival probabilities at clinically meaningful landmarks, or the probability of achieving a prespecified clinically meaningful benefit. A Bayesian approach would offer several advantages over a deterministic projection based on a single early survival landmark. It would naturally provide uncertainty quantification for predicted medians. It could formally incorporate prior information and allow dynamic updating as follow-up matures. Moreover, posterior predictive probabilities could support interim monitoring and decision making in a manner aligned with clinical questions. In January 2026, the US Food and Drug Administration issued a draft guidance titled Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products.11 The guidance addresses the use of Bayesian methods in clinical trials, including prior specification, decision criteria, operating characteristics, and primary inference. Perspectives on the application of Bayesian methods in clinical trials have also been published.12-14 These developments make the consideration of Bayesian prediction methods for immature survival data particularly timely. The work by Opperman et al. addresses an important and practical problem: how to estimate clinically meaningful survival outcomes before the median is observed. Their method is potentially useful for early interpretation of emerging survival data. At the same time, its limitations highlight the need for uncertainty quantification, careful consideration of nonconstant hazards, endpoint-specific features, and treatment classes. Future work should evaluate methods that reconstruct individual-level survival data from KM curves, quantify uncertainty in projecting future outcomes, compare median-based predictions with alternative estimands such as RMST, and use Bayesian models to incorporate prior information and generate posterior predictive distributions from mature outcomes. As the saying commonly attributed to Niels Bohr reminds us, “Prediction is very difficult, especially about the future.”15 This work was partially supported by the grants from the National Cancer Institute and the John G. & Marie Stella Kenedy Foundation Chair in Cancer Research at The University of Texas MD Anderson Cancer Center. The funders had no role in the design of the study; the collection, analysis, or interpretation of the data; or the writing of the manuscript and decision to submit it for publication. J. Jack Lee (Conceptualization, Writing—original draft, Writing—review & editing) J.J.L.'s research was supported in part by the grants P30CA016672, P01CA285249, U24CA224285, and U24CA274274 from the National Cancer Institute and the John G. & Marie Stella Kenedy Foundation Chair in Cancer Research. No new data were generated or analyzed for this editorial.