Learning strategies for individualized treatment rules

PhD student: Etienne Peyrot

Title: Learning strategies for individualized treatment rules

Supervisor: François Petit

Doctoral school: ED 393 Epidemiology and Biomedical Information Sciences, Université Paris Cité

Thesis topic:

Scientific background to the project One of the aims of personalized medicine is to offer a patient the most appropriate treatment based on his or her characteristics. In other words, this means determining the optimal therapy for each individual, i.e. the one that maximizes the benefit/risk ratio. One of the most common approaches is to consider only the benefits of treatment and to use statistical learning methods (machine learning) to construct an individual treatment rule, which recommends for each individual the treatment that will have the greatest effect. The construction of these rules generally involves learning the individual treatment effect function, obtained as the difference of the expected judgment criterion under one of the compared treatments, conditional on a set of covariates representing the characteristics of the individuals that will be used to individualize the treatment recommendation. This type of approach requires the estimation of a large number of nuisance parameters, such as the effect of covariates related to prognosis (judgment criterion, but not to a differential treatment effect). Recent work shows that it is possible, by transforming the initial problem into a continuous minimization problem, to directly learn a treatment benefit score, without estimating these nuisance parameters. The magnitude of this score also provides information on the magnitude of the individualized treatment effect, from which the optimal treatment rule is immediately deduced. However, this approach requires the ability to construct loss functions with specific properties.

If we consider that the aim of individualized treatment rules is to maximize benefit and minimize risk, not just maximize benefit, the situation is more complex. This is a multi-objective optimization problem. Moreover, much of the literature on the construction of these rules systematically includes risk as a constraint, and is therefore limited to the ε-constraint method of multiobjective optimization, whereas there is a wide range of methods for multiobjective optimization. Finally, it is desirable to be able to construct these treatment rules 1) from randomized trial data as well as observational data, 2) with potentially smaller sample sizes.

This thesis project focuses on methods for constructing individualized treatment rules that take into account the risk/benefit ratio of treatment. It is divided into three phases.

1. The first step is to adapt the general approach to the case of benefit/risk rule construction, and to study and identify the relevant multi-objective optimization methods for this rule construction problem.
2. We want to be able to generate treatment rules from both randomized trial data and observational data, which may be small in size. In recent years, several new balancing methods have emerged. Some of them look very promising. We will be evaluating them intensively. This will enable us to identify the most suitable for smaller sample sizes.
3. The final step is to adapt and implement the results of steps 1 and 2 on small-scale observational data. The method identified in step 2 as the most efficient on small samples will be used to perform the required balancing.