Desirability of outcome ranking (DOOR) analysis for multivariate survival outcomes with application to ACTT-1 trial

TL;DR

This paper introduces a novel statistical approach combining DOOR analysis with RMST to better evaluate benefit-risk in clinical trials, accounting for competing risks and patient trajectories.

Contribution

It develops estimation and inference procedures that integrate DOOR and RMST, providing more intuitive and comprehensive analysis of multivariate survival outcomes.

Findings

Proposed a multivariate Gaussian process estimator for RMSTs.

Validated methods through simulations under various scenarios.

Applied approach to ACTT-1 trial data.

Abstract

Desirability Of Outcome Ranking (DOOR) methodology accounts for problems that conventional benefit:risk analyses in clinical trials ignore, such as competing risks and the trade-off relationship between efficacy and toxicity. DOOR levels can be considered as a multi-state process in nature, as event-free survival, and survival with side effects are not equivalent and the overall patient trajectory requires recognition. In monotone settings where patients' conditions can only decline, we can record event times for each transition from one level of the DOOR to another, and construct Kaplan-Meier curves displaying transition times. While traditional survival analysis methods such as the Cox model require assumptions like proportional hazards and suffer from the challenge of interpreting a hazard ratio, Restricted Mean Survival Time (RMST) offers an alternative with greater intuitiveness. Therefore, we propose a combination of the two domains to develop estimation and inferential procedures that could benefit from the advantages of both DOOR and RMST. Particularly, the area under each survival curve restricted to a time point, or the RMST, has clear clinical meanings, from expected event-free survival time, expected survival time with at most one of the events, to expected lifetime before death. We show that the nonparametric estimator of the RMSTs asymptotically follows a multivariate Gaussian process through the martingale theory and functional delta method. There are alternative approaches to hypothesis testing that recognize when patients transition into worse states. We evaluate our proposed method with data simulated under a multistate model. We consider various scenarios, including when the null hypothesis is true, when the treatment difference exists only in certain DOOR levels, and small-sample studies. We also present a real-world example with ACTT-1.