Dec 14, 2018 | 16:45—17:30
Multimorbidity is a major health issue, especially in the elderly, and better understanding the complex higher-order interactions between diseases is a goal of this work. In this seminar talk, I present our machine learning based approach of building an interpretable predictive model to characterize the multimorbid health state of patients. The non-parametric model allows in particular the quantification of synergistic/redundant associations between disease phenotypes, or, eventually the identification of new disease phenotypes. In this work, a disease phenotype is technically a set of diagnoses, but these phenotypes (features) gain meaning by their relation to (single) diagnoses (targets) via the coefficients of the predictive model. These model coefficients form a multiplex network where the phenotypes are the nodes and the targets make up the layers. The phenotypes within a layer form a lattice structure which allows to relate higher-order phenotypes to their lower-order phenotype neighbors. The multiplex network can be reduced to a target-target or a phenotype-phenotype network. By extracting the backbones of these networks, potentially interesting relations between diseases can be deduced. The presentation will explain the method and show some preliminary findings.