A Novel R Shiny Tool TVCurveTM for Survival Analysis with Time-Varying Covariate in Oncology Clinical Studies: Overcoming Biases and Enhancing Collaboration

Abstract

In oncology studies, analyzing survival outcomes with time-varying covariates (TVCs), like the impact of receiving hematopoietic cell transplantation (HCT) after chimeric antigen receptor T cell (CAR-T) infusion, faces challenges for standard methods like Cox models and Kaplan-Meier curves, which assume TVC status is fixed at baseline, causing biased estimates. Landmark analysis is an alternative but is limited by its dependence on a chosen start time. Time-dependent (TD) Cox model is better suited for TVC analysis, though its visualization can be complex. A novel visualization, Smith-Zee, based on a TD Cox model, addresses this issue by mimicking new patients with TVC status change at different times, which addresses drawbacks of the Naïve and Landmark methods. However, software to implement these methods is limited.

To address this, we developed TVCurve^TM, an R Shiny tool that integrates various models (Naïve, landmark, and TD Cox) and curves (Naïve KM, Landmark KM, Smith-Zee, and Extended KM). https://samplen.shinyapps.io/TVCurve/.

Using two CAR-T trials as examples (Fred Hutch: NCT01865617, Seattle Children's: NCT02028455), we re-analyzed the impact of HCT on leukemia-free survival (LFS) by applying the Naïve, landmark, and TD Cox models. We observed that the Naïve and TD Cox yielded similar hazard ratios (HRs) and p-values (Table 1), indicating consistent results from both studies. However, the choice of landmark time leads to varying HRs and p-values (Table 1), highlighting the sensitivity of the landmark approach.

To understand and demonstrate the limitations of the Naïve method, we conducted a simulation study on TVCurve^TM. Our results revealed increased bias when more patients experience late TVC changes, and minimal bias when more patients had earlier TVC changes. This aligns with the findings from our analysis of two CAR-T studies, where the administration of hematopoietic cell transplantation (HCT) occurred promptly after CAR-T infusion and a substantial majority (90%) of patients exhibited treatment response, leading to longer RFS. To visualize the impact of different parameter combinations on the bias, we provided Contour plots in the simulation panel on TVCurve^TM.

In conclusion, the TD Cox model and Smith-Zee curves are recommended for survival analysis with TVCs. More importantly, our TVCurve^TM breaks collaboration barriers since it does not require data sharing but ensure standardized analyses across datasets, offering a vital tool for advancing survival analysis in oncology research.