Abstract
One of the targets of personalized medicine is to provide treatment recommendations using patient characteristics. We present the command ptr, which both predicts a personalized treatment recommendation algorithm and evaluates its effectiveness versus an alternative regime, using randomized trial data. The command allows for multiple (continuous or categorical) biomarkers and a binary or continuous outcome. Confidence intervals for the evaluation parameter are provided using bootstrap resampling.
Highlights
One of the goals of the modern paradigm of “personalized medicine” is to use data collected on a patient to recommend treatment, rather than treating all patients with the same therapy
In the following two sections, we describe two estimates of θ: a nonparametric inverse probability weighted (IPW) estimate and the more efficient augmented inverse probability weighted (AIPW) estimate
We explained how a personalized treatment recommendation (PTR) can be estimated and evaluated using data retrospectively collected from two-armed randomized controlled trials. We implemented this theory in a package: ptr
Summary
One of the goals of the modern paradigm of “personalized medicine” is to use data collected on a patient to recommend treatment, rather than treating all patients with the same therapy. Estimating a PTR differs from diagnostic or prognostic model development because the object of the inference (whether a patient benefited from treatment) remains strictly unobserved. Once a PTR has been estimated, it is of interest to know whether it improves on a conservative policy, for example, one where everybody is treated. This potential improvement can be estimated prospectively using a new trial, or it can be done retrospectively using a separate dataset from the one used to develop the model. We demonstrate the community-contributed ptr command to first estimate a PTR using regression modeling and to separately estimate the benefit of a PTR compared with an alternative policy, using retrospectively collected data. The command is implemented in the context of a two-armed, randomized controlled trial with a single treatment decision, a binary treatment, and continuous or binary outcomes
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have