Abstract:
|
In this talk, we discuss recent developments in machine learning for advancing the precision medicine quest. One such development, outcome weighted learning, directly estimates optimal treatment rules without modeling the primary outcome as a function of patient-level features. The new approach helps leverage treatment heterogeneity to discover treatment rules with complex and multi-stage treatment options, including options on a continuum such as dose level or administration timing. We propose a clinical trial design where candidate dose levels are randomly assigned from a continuous distribution within a safe range. An outcome weighted learning method, based on a non-convex loss function, is used to efficiently estimate the individualized dose rule (IDR) via a difference of convex functions algorithm. Consistency and convergence rates for the estimated IDR are derived, and the approach is evaluated via simulation studies and an analysis of data from a cohort study for Warfarin (an anti-thrombotic drug) dosing. The new approaches expand the use of machine learning technology in precision medicine.
|