Trustworthy Bandit Algorithms for Precision Medicine

Contextual bandit algorithms learn personalized recommendation policies using data on individual responses to past recommendations.

These machine learning methods are therefore particularly relevant for precision medicine, where the goal is to learn treatment recommendations adapted to patient characteristics so as to improve the medical benefit in a population.

In this setting, bandit algorithms operate offline by learning recommendation policies using randomized trial or observational data.

This raises fundamental challenges which cannot be overcome by increased amounts of data, and is the focus of this project.For learned treatment recommendation policies to be trustworthy and adhere to the cautionary principle of “above all, do no harm”, the learning methods require valid certification of health performance and trade-offs.

Conventional methods can easily learn erroneous recommendations with invalid performance certifications, since there are always unmeasured factors that jointly affect health outcomes and the selection of individuals for trials or past clinical decisions.

This project will tackle these fundamental challenges to learning policies from randomized trial or observational data by developing new learning theory and trustworthy methods for precision medicine.

It will draw on advances in machine learning and causal inference, using nonparametric techniques and bounds on miscalibrated models to certify policy learning.