PhD student: Florie Bouvier
Title: Estimation of individualized treatment effects using an individual participant data meta-analysis (temporary title)
Supervisors: Raphaël Porcher
Doctoral school: ED 393 Epidemiology and Biomedical Information Sciences, Université Paris Cité
Thesis topic:
Personalized medicine aims at tailoring treatment to the individual characteristics of each patient. One of the central aspects of personalized medicine is identifying subgroups of patients who benefit from one intervention or more from one intervention than another. Several approaches exist to determine these subgroups, particularly the individualized treatment effect (ITE) estimation. The ITE represents the expected treatment effect for an individual given its characteristics. In this thesis, we aimed to study the development of personalized treatment rules by estimating individualized treatment effects, based on data from one or more randomized clinical trials. We first sought to estimate ITEs using an individual participant data meta-analysis. We then compared the treatment rules produced by various methods. Finally, we studied the maximum discrimination that could be obtained for various distributions of treatment effects.
In general, predictive models for ITE are based on data from single randomized controlled trials (RCTs), which are not always large enough to achieve this objective, leading to a risk of overfitting the models or, on the contrary, the impossibility of capturing the effect of relevant variables. Individual participant data meta-analyses (IPD-MA) offer a solution by pooling data from several RCTs, thus potentially improving the generalizability of results. Our first project explored the estimation of ITEs via an IPD-MA, evaluating the performance of different methods that consider heterogeneity between studies in the meta-analysis, combined with two methods for estimating ITE (S-learner and T-learner). The results show that integrating interactions with treatment (S-learner) is beneficial, without any specific method standing out in terms of performance.
Identifying subgroups that benefit from a specific intervention makes it possible to develop individualized treatment rules (ITR). Numerous machine-learning techniques have recently been proposed to generate these rules. Nevertheless, the consistency of the ITRs produced by these methods, i.e. whether they recommend the same treatment to the same individuals, remains uncertain. Our second project compared the ITRs generated by 22 methods in two clinical trials, revealing that the methods result in different ITRs and are therefore not interchangeable. The choice of method greatly influences treatment recommendations for patients, raising concerns about their practical use.
To develop effective individualized treatment rules, these rules must have a good capacity to distinguish patients who benefit from taking a certain treatment from patients who do not. In other words, these rules need to have good discrimination. In my last project, we explored the maximum level of discrimination that could be achieved for different distributions of treatment effects, each with a different level of heterogeneity. We selected three discrimination metrics: the c-statistic for benefit, the Concentration of Benefit, and the Population Average Prescription Effect (PAPE). Our results indicate that the presence of heterogeneous treatment effects is not systematically translated into favorable discrimination results. Optimal discrimination depends on the distribution of treatment effects. Furthermore, the choice of metric used to assess discrimination impacts the conclusions, as they do not require the same levels of treatment effect heterogeneity to lead to favorable discrimination.
Keywords: personalized medicine, individualized treatment effect, heterogeneity of treatment effect, individualized treatment rule, prediction model, machine learning, individual participant data meta-analysis
