Lerning from signals: statisical learning methods for adverse reaction detection and functional data clustering

This thesis addresses two research questions motivated by the practical needs of Arithmos Srl. The first concerns the identification of potential Adverse Drug Reactions (ADRs) in marketed drugs. To tackle this problem, a Random Forest model was developed and evaluated on a dataset of interest to the company. The results suggest that this approach improves ADR detection compared with the disproportionality measures commonly used in pharmacovigilance practice. The second research question involves the development of clustering tools for functional data. Our approach builds on the Random Projections framework, which simultaneously reduces dimensionality and provides multiple representations of the data, leading to diverse clustering solutions that are later combined through an ensemble procedure. Based on these ideas, we implemented a model whose performance was assessed through both simulations and real data applications, showing promising results. Finally, we present preliminary work on a multivariate extension of the one-dimensional functional model. Although still under development, the early results indicate that the proposed approach is promising and merits further refinement.