Title: A hybrid approach in dynamic treatment regimes problems: Multivariate Bayesian machine learning
Authors: Edoardo Ghezzi - University of Milano-Bicocca (Italy) [presenting]
Matteo Borrotti - University of Milan-Bicocca (Italy)
Abstract: A dynamic treatment regime (DTR) is a sequence of decision rules, one per stage of intervention, which aims to adapt a treatment plan to the time-varying state of an individual subject. The Bayesian Machine Learning approach (BML) avoids many of the problems arising from the use of other common methods, such as Q-learning, for identifying optimal DTRs. In problems concerning personalized medicine, it is often plausible to have a large set of variables. In such scenarios, the BML approach might be incalculable due to the singularity of the design matrix and therefore, its unfeasible reversibility. The suggested approach, known as Multivariate Bayesian Machine Learning (MBML), consists of an initial variable selection performed by Spike and Slab priors, in particular the independence slab (i-slab), followed by a generalization of the classic BML approach, which enables its use in a multivariate scenario. Through a simulation study, the MBML approach is compared with Q-learning in various two-stage settings; these settings differ in complexity, as they have several sample sizes and dimensionalities, and in regularity conditions, as they vary each stage's coefficients. The MBML approach has given proof of its reliability by showing results which were more accurate than those of Q-learning in almost every scenario.