A flexible threshold biomarker trial design for treatment intensification and de-intensification
Weilin Wang, Kelley M. Kidwell, Robert T. Dess, Matthew J. SchipperBackground:
In Oncology, prognostic biomarkers are commonly used to tailor treatment intensity to a patient’s risk of disease progression. Several recent Phase III trials embed two studies within one protocol, comparing standard versus de-intensified therapy in good-prognosis patients and intensified versus standard therapy in poor-prognosis patients, using a single fixed biomarker threshold to classify patients into good and poor prognosis. Such designs are clinically intuitive but rely on a prespecified cut-point for which there can be substantial uncertainty as to the optimal value.
Methods:
We propose a flexible threshold, biomarker-based design in which patients with intermediate biomarker values are randomized to three arms (control, de-intensified, intensified), while those at the extremes are randomized between control and de-intensified therapy (low biomarker) or intensified (high biomarker). The design jointly tests non-inferiority of de-intensified therapy in lower-risk patients and superiority of intensified therapy in higher-risk patients, while estimating the optimal biomarker thresholds. Family-wise Type I error is controlled through a two-stage bootstrap testing procedure.
Results:
In simulation settings where the alternative non-inferiority and superiority hypotheses hold over broad overlapping biomarker regions, the proposed flexible threshold biomarker design has greater power than a fixed threshold design while maintaining appropriate Type I error control. When the optimal decision rule matches the prespecified fixed cutoff, the two designs perform similarly. This reflects a bias–variance trade-off, in which fixed threshold designs perform well when the cutoff is correctly specified, whereas the flexible threshold biomarker design is more robust when the optimal threshold is uncertain or misspecified.
Conclusions:
The proposed flexible threshold biomarker design offers a robust and efficient framework for trials seeking to tailor treatment intensity using a prognostic biomarker. The approach is broadly applicable to other diseases in which both escalation and de-escalation of therapy are of clinical interest and there is clinical equipoise among the three treatment options for patients with intermediate biomarker values.