Biomarker-guided adaptive enrichment design with threshold detection for clinical trials with time-to-event outcome

Kaiyuan Hua, Hwanhee Hong, Xiaofei Wang

TL;DR

This paper introduces a novel two-stage clinical trial design that adaptively identifies the best biomarker threshold to "enrich" a study with patients most likely to benefit from a treatment. By using Restricted Mean Survival Time (RMST) instead of traditional hazard ratios, the design remains robust even when treatment effects are delayed or non-proportional.

What Problem Does This Paper Address?

In the era of precision medicine, researchers often use biomarkers (like protein expression levels) to identify which patients will respond to a new drug. However, choosing the "right" biomarker cutoff before a trial starts is a gamble; if you set the threshold too high or too low, you might miss the treatment effect or include too many non-responders, diluting the results.

Furthermore, most survival-based trials rely on the Hazard Ratio (HR). The HR assumes "proportional hazards"—the idea that the relative risk of an event between two groups stays constant over time. In modern immunotherapy trials, this assumption is frequently violated (e.g., a drug takes months to start working, creating a "delayed effect"). When this happens, the HR becomes a misleading and underpowered metric.

Methods

The authors propose a two-stage biomarker-guided adaptive RMST design:

  1. Stage I (Exploration): Patients are recruited regardless of their biomarker status. At the end of this stage, the researchers use a Prediction Method (based on piecewise exponential models) to estimate a biomarker cutpoint ($c_{t^*}$) where the treatment starts showing a positive effect.
  2. Stage II (Enrichment): If a clear benefit-subgroup is identified, enrollment is restricted only to "biomarker-positive" patients.
  3. Final Analysis: To measure the treatment effect, the authors use Restricted Mean Survival Time (RMST)—essentially the area under the survival curve up to a specific time point. This represents the average "gain" in survival time (e.g., "patients on Drug A lived 3 months longer on average than Drug B").

The "Secret Sauce": Calibration Weighting (CW) Because Stage II only recruits "positive" patients, the final pool of patients has a different biomarker distribution than the original population. To fix this "distributional shift" and prevent biased results, the authors introduce Calibration Weighted Estimators. These mathematical tools re-weight the data so the final estimate reflects the true treatment effect in the target population.

Datasets

As a methodology paper, the authors do not introduce a new massive public dataset. Instead, they evaluate their approach using:

  • Numerical Re-design: They used parameters from the JAVELIN Lung 200 oncology trial (avelumab vs. docetaxel) to demonstrate how their design would have performed in a real-world scenario.
  • Extensive Simulations: 10,000 simulated datasets were generated to test the design under various conditions, such as different hazard functions and varying levels of prior information about "change points" in survival.

Key Insights and Ideas

  • RMST > Hazard Ratio: RMST is more intuitive for clinicians and patients (it talks about time gained, not abstract ratios) and doesn't break when the drug has a delayed effect.
  • Don't Guess, Adapt: By identifying the threshold during the trial, researchers avoid the "all-or-nothing" risk of pre-specifying an incorrect biomarker cutoff.
  • Efficiency Gains: The "Enrichment" strategy significantly boosts statistical power. By focusing on the responders, you can achieve the same level of scientific certainty with fewer total patients.
  • The G-Formula Winner: Among the five estimators tested, the CW G-Formula Estimator consistently showed the highest statistical power and the smallest variance, making it the recommended choice for these types of trials.

Key Results

  • Sample Size Reduction: In the JAVELIN Lung 200 re-design, the proposed enrichment design required only 845 patients to reach 90% power, compared to 940 patients required for a traditional "all-comer" design. That’s a ~10% reduction in required participants.
  • Bias Correction: The simulations proved that "Naive" estimators (which ignore the enrichment shift) overestimate treatment effects. The proposed Calibration Weighted (CW) estimators successfully removed this bias, maintaining a 95% coverage probability.
  • Error Control: The design effectively controlled the "family-wise type I error rate," ensuring that the probability of a "false positive" result remained below the strict 2.5% threshold.

Limitations and Future Work

  • Single Biomarker Focus: Currently, the model only handles one continuous biomarker. Real-world response is often driven by a complex "signature" of multiple biomarkers.
  • Truncation Time Sensitivity: RMST requires choosing a "truncation time" ($t^*$). If the trial doesn't follow patients long enough to reach this point, the method becomes difficult to implement.
  • Piecewise Assumptions: The early-stage prediction method relies on piecewise exponential hazard assumptions. While flexible, it still requires some prior knowledge of when treatment effects might change.
  • Future Directions: The authors suggest extending this to "group-sequential" designs (allowing for even earlier stopping for success or failure) and expanding the math to handle multiple biomarkers simultaneously.

Summary updated

Abstract

Biomarker-guided designs are increasingly used to evaluate personalized treatments based on patients' biomarker status in Phase II and III clinical trials. With adaptive enrichment, these designs can improve the efficiency of evaluating the treatment effect in biomarker-positive patients by increasing their proportion in the randomized trial. While time-to-event outcomes are often used as the primary endpoint to measure treatment effects for a new therapy in severe diseases like cancer and cardiovascular diseases, there is limited research on biomarker-guided adaptive enrichment trials in this context. Such trials almost always adopt hazard ratio methods for statistical measurement of treatment effects. In contrast, restricted mean survival time (RMST) has gained popularity for analyzing time-to-event outcomes because it offers more straightforward interpretations of treatment effects and does not require the proportional hazard assumption. This paper proposes a two-stage biomarker-guided adaptive RMST design with threshold detection and patient enrichment. We develop sophisticated methods for identifying the optimal biomarker threshold and biomarker-positive subgroup, treatment effect estimators, and approaches for type I error rate, power analysis, and sample size calculation. We present a numerical example of re-designing an oncology trial. An extensive simulation study is conducted to evaluate the performance of the proposed design.
BibTeX
@article{hua2025biomarker,
  author = {Kaiyuan Hua and Hwanhee Hong and Xiaofei Wang},
  title = {{Biomarker-guided adaptive enrichment design with threshold detection for clinical trials with time-to-event outcome}},
  journal = {Journal of Biopharmaceutical Statistics},
  year = {2025},
  doi = {10.1080/10543406.2025.2489291},
}

Loading related content...

Figure