Optimal individualized treatment regimes for survival data with competing risks

Christina W. Zhou, Nikki L. B. Freeman, Katharine L. McGinigle, Michael R. KosorokarXiv 2024 · DOI · arXiv

This paper, recently published by researchers from UNC-Chapel Hill and Duke University, introduces a novel framework for determining the best medical treatment for a patient when multiple, mutually exclusive risks are at play. By combining survival analysis with a "multi-utility" decision process, the authors provide a tool that helps clinicians balance the goal of keeping a patient alive with the goal of avoiding specific devastating complications.

TL;DR

The authors develop a nonparametric, two-phase estimator for Individualized Treatment Regimes (ITRs) that accounts for "competing risks" (e.g., death vs. limb loss). By using a novel multi-utility value function and random forest-based estimators, the method prioritizes overall survival while minimizing the risk of a specific "priority" event when survival outcomes are comparable.

What Problem Does This Paper Address?

In precision medicine, the goal is to use data to recommend the best treatment for a specific individual. However, in complex diseases like Peripheral Artery Disease (PAD), patients face competing risks. For example, a patient might be at risk for both death and major amputation; if the patient dies, they are no longer at risk for amputation.

Standard statistical methods often struggle here. They either focus solely on "how long will the patient live?" (Overall Survival) or "what is the risk of this specific complication?" (Cumulative Incidence). No existing method simultaneously optimizes for both in a way that reflects clinical reality—where a doctor might be willing to accept a tiny, statistically insignificant decrease in survival if it significantly lowers the risk of a life-altering complication like losing a limb.

Methods

The core of the paper is a two-phase, multi-utility optimization strategy:

  1. Phase 1 (Survival): The model first identifies treatments that maximize Overall Survival (OS). It uses a Random Survival Forest (RSF) to estimate individual survival curves.
  2. Phase 2 (Risk Mitigation): If multiple treatments yield survival results within a user-defined "tolerance level" ($\alpha_\phi$), the model moves to Phase 2. Here, it chooses the treatment that minimizes the Cumulative Incidence Function (CIF) of a "Priority Cause" (the specific complication of interest).
  3. Random Cumulative Incidence Forest (RCIF): To power Phase 2, the authors developed a new forest-based estimator. Unlike standard forests, this one uses Gray’s test—a specialized statistical test for competing risks—as the splitting criterion for the trees.

This approach is nonparametric, meaning it doesn't make rigid assumptions about the shape of the data, allowing it to capture complex, non-linear interactions between patient characteristics and treatment outcomes.

Datasets

The authors validated their method using:

  • Synthetic Simulations: Two different data-generating mechanisms (Exponential and Fine-Gray settings) were used to test the model's consistency and performance across varying levels of censoring and covariate complexity.
  • CLTI Cohort: A real-world dataset of 303 patients with Chronic Limb-Threatening Ischemia (CLTI) from the North Carolina Translational and Clinical Sciences Institute. The "Priority Cause" was death, with major amputation treated as the competing risk.

Key Insights and Ideas

  • Hierarchical Decision Making: Instead of trying to combine survival and risk into a single, confusing "score," the authors use a hierarchical approach that mirrors clinical logic: "First, keep them alive; then, if survival is similar, pick the treatment with the fewest side effects."
  • The Power of Tolerance ($\alpha_\phi$): The inclusion of a tolerance parameter allows clinicians to define what "similar survival" means in practice. This makes the model a flexible tool rather than a black box.
  • Forests for Competing Risks: By adapting Gray’s test for node splitting in a random forest, the authors created a more robust way to estimate individualized risk curves than traditional linear models.

Key Results

  • Simulation Performance: In synthetic tests, the proposed method (implemented in the R package itrSurv) showed lower variance and higher average survival compared to existing methods like AIPWE or PMCR.
  • Clinical Application: When applied to the CLTI cohort, the model successfully identified treatment regimes. Interestingly, the "zero-order" version (recommending the same treatment for everyone) performed very well, suggesting that in some datasets, the overall treatment effect is so strong it outweighs individual heterogeneity.
  • Balanced Outcomes: Compared to models that only look at one metric, this method consistently found a "middle ground"—achieving high survival without ignoring the specific risks of the priority cause.

Limitations and Future Work

  • Single Priority Cause: The current model assumes there is one "main" complication to avoid. It does not yet handle scenarios with three or more equally important competing risks.
  • Tie-Breaking: The authors note that the model lacks a formal tie-breaker if two treatments are identical across both survival and priority risk.
  • Single-Stage Decisions: The current framework is designed for a single treatment decision point. The authors suggest that extending this to "multi-stage" regimes (where treatments change over time) is a promising area for future research.
  • Data Sensitivity: Like all forest-based methods, performance is highly dependent on having a large enough sample size to distinguish between patient subgroups, which can be a challenge in rare disease cohorts.

Software Availability: The authors have made their R package available at https://github.com/cwzhou/itrSurv.

Summary updated

Abstract

Precision medicine leverages patient heterogeneity to estimate individualized treatment regimens, formalized, data-driven approaches designed to match patients with optimal treatments. In the presence of competing events, where multiple causes of failure can occur and one cause precludes others, it is crucial to assess the risk of the specific outcome of interest, such as one type of failure over another. This helps clinicians tailor interventions based on the factors driving that particular cause, leading to more precise treatment strategies. Currently, no precision medicine methods simultaneously account for both survival and competing risk endpoints. To address this gap, we develop a nonparametric individualized treatment regime estimator. Our two-phase method accounts for both overall survival from all events as well as the cumulative incidence of a main event of interest. Additionally, we introduce a multi-utility value function that incorporates both outcomes. We develop random survival and random cumulative incidence forests to construct individual survival and cumulative incidence curves. Simulation studies demonstrated that our proposed method performs well, which we applied to a cohort of peripheral artery disease patients at high risk for limb loss and mortality.
BibTeX
@misc{zhou2024optimal,
  author = {Christina W. Zhou and Nikki L. B. Freeman and Katharine L. McGinigle and Michael R. Kosorok},
  title = {{Optimal individualized treatment regimes for survival data with competing risks}},
  year = {2024},
  eprint = {2411.08315},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.08315},
}
Figure