Transformation-Invariant Learning of Optimal Individualized Decision Rules with Time-to-Event Outcomes

Yu Zhou, Lan Wang, Rui Song, Tuoyi Zhao

Abstract

In many applications of precision medicine, the primary goal is to identify an optimal individualized decision rule (IDR) to prolong survival time. However, survival data often present challenges such as censoring and heavy-tailed distributions. Existing methods typically focus on maximizing the restricted mean survival time, which often requires restrictive model assumptions. We propose a robust framework for estimating optimal static or dynamic IDRs based on a quantile criterion. This approach is transformation-invariant, does not require an outcome regression model, and is robust to outliers. We establish the statistical theory for the estimated parameters, prove consistency even under the challenge of non-unique optimal rules (exceptional laws), and provide a smoothed resampling procedure for inference. The methods are implemented in the R-package QTOCen.

Introduction

The core objective of precision medicine is to develop decision rules that recommend the most effective treatment for a patient based on their unique clinical and demographic profile. When the outcome of interest is time-to-event—such as time to relapse or death—researchers face two significant hurdles: censoring and the sensitivity of mean-based metrics to skewed data.

Existing methodologies primarily focus on maximizing the Restricted Mean Survival Time (RMST). While useful, RMST-based approaches often rely on restrictive parametric or semi-parametric models. If these models fail to capture the underlying data-generating process, the resulting IDR can be biased. This paper addresses these limitations by shifting the focus to a quantile-based optimality criterion, offering a more robust and flexible alternative for clinical decision-making.

Methodology

The proposed framework seeks to find a decision rule that maximizes the $\tau$-th quantile (e.g., the median) of the potential survival time distribution.

Quantile Optimality and Invariance

By optimizing for quantiles rather than the mean, the resulting IDR becomes transformation-invariant. This ensures that the optimal treatment recommendation remains the same under any strictly increasing transformation of the survival time (e.g., measuring survival in days versus months, or using a logarithmic scale). This property is a significant advantage over mean-based methods, which can yield different recommendations depending on the data scale.

Handling Censored Data

To account for right-censoring, the authors utilize Inverse Probability Weighting (IPW). By weighting the observed outcomes using Kaplan-Meier estimators of the censoring distribution, the framework provides unbiased estimates of the quantile-based value function without requiring a pre-specified regression model for the survival time. This "model-free" nature protects against bias from model misspecification.

Optimization and Inference

The objective function for learning the IDR is non-smooth due to the presence of indicator functions (e.g., "if a covariate exceeds a threshold, then treat"). To find the global maximum, the authors employ a genetic algorithm via the rgenoud package. Furthermore, because standard bootstrap methods often fail for non-regular estimation problems with $n^{1/3}$ convergence rates, a novel smoothed resampling procedure is introduced to facilitate reliable statistical inference and confidence interval construction.

Empirical Evaluation

The framework was evaluated using both synthetic simulations and real-world clinical data.

Simulation Studies

Extensive Monte Carlo studies demonstrated that the quantile-based approach significantly outperforms naive estimators, particularly in the presence of heavy-tailed survival distributions and complex censoring mechanisms. The method also proved effective in dynamic, multi-stage decision settings where the second treatment stage depends on the patient's initial response, maintaining high accuracy even with 40% censoring.

Application to Breast Cancer Data

The authors applied their method to the German Breast Cancer Study Group (GBSG2) dataset, which includes 686 patients with node-positive breast cancer. Key findings included:

  • Hormone Receptor Levels: Estrogen Receptor (ER) and Progesterone Receptor (PR) levels were identified as the most critical factors in determining the benefit of Tamoxifen therapy.
  • Clinical Interpretability: For certain survival quantiles, the framework suggested that age was not a significant factor, allowing for the creation of more concise and actionable clinical rules.

Conclusion

The proposed quantile-based framework provides a robust, transformation-invariant alternative to mean-based IDR estimation in survival analysis. By avoiding the need for outcome regression models and handling censoring through IPW, it offers a flexible tool for precision medicine. The implementation in the QTOCen R-package allows practitioners to apply these advanced statistical techniques to optimize treatment strategies in real-world clinical trials.

Summary updated

Abstract

In many important applications of precision medicine, the outcome of interest is time to an event (e.g., death, relapse of disease) and the primary goal is to identify the optimal individualized decision rule (IDR) to prolong survival time. Existing work in this area have been mostly focused on estimating the optimal IDR to maximize the restricted mean survival time in the population. We propose a new robust framework for estimating an optimal static or dynamic IDR with time-to-event outcomes based on an easy-to-interpret quantile criterion. The new method does not need to specify an outcome regression model and is robust for heavy-tailed distribution. The estimation problem corresponds to a nonregular M-estimation problem with both finite and infinite-dimensional nuisance parameters. Employing advanced empirical process techniques, we establish the statistical theory of the estimated parameter indexing the optimal IDR. Furthermore, we prove a novel result that the proposed approach can consistently estimate the optimal value function under mild conditions even when the optimal IDR is nonunique, which happens in the challenging setting of exceptional laws. We also propose a smoothed resampling procedure for inference. The proposed methods are implemented in the R-package QTOCen. We demonstrate the performance of the proposed new methods via extensive Monte Carlo studies and a real data application. Supplementary materials for this article are available online.
BibTeX
@article{zhou2022transformation,
  author = {Yu Zhou and Lan Wang and Rui Song and Tuoyi Zhao},
  title = {{Transformation-Invariant Learning of Optimal Individualized Decision Rules with Time-to-Event Outcomes}},
  journal = {Journal of the American Statistical Association},
  year = {2022},
  doi = {10.1080/01621459.2022.2068420},
}
Figure