Gaussian Processes for Individualized Continuous Treatment Rule Estimation

TL;DR

This paper introduces a Gaussian process-based method for estimating individualized continuous treatment rules, explicitly modeling outcome uncertainty to improve treatment selection and outperform existing approaches.

Contribution

It presents a novel Gaussian process regression approach that directly incorporates outcome uncertainty into continuous ITR estimation, enhancing treatment decision quality.

Findings

Outperforms standard methods in value function maximization

Provides better interpretability of treatment rules

Easier to extend to multiple continuous treatments

Abstract

Individualized treatment rule (ITR) recommends treatment on the basis of individual patient characteristics and the previous history of applied treatments and their outcomes. Despite the fact there are many ways to estimate ITR with binary treatment, algorithms for continuous treatment have only just started to emerge. We propose a novel approach to continuous ITR estimation based on explicit modelling of uncertainty in the subject's outcome as well as direct estimation of the mean outcome using gaussian process regression. Our method incorporates two intuitively appealing properties - it is more inclined to give a treatment with the outcome of higher expected value and lower variance. Experiments show that this direct incorporation of the uncertainty into ITR estimation process allows to select better treatment than standard indirect approach that just models the average. Compared to the competitors (including OWL), the proposed method shows improved performance in terms of value function maximization, has better interpretability, and could be easier generalized to multiple interdependent continuous treatments setting.