A0523
Title: Flexible parametrization of graph-theoretical features from individual-specific networks for prediction
Authors: Mariella Gregorich - Medical University of Vienna (Austria) [presenting]
Abstract: Nowadays, many structurally similar predictors are available for each individual, which can be represented as individual-specific networks able to capture their dependence structure and can provide predictive biomarkers for outcome modelling. However, unsubstantiated, arbitrary decisions in individual-specific network inference, in particular when choosing a suitable threshold for network sparsification, still lead to a high variability of the extracted graph-theoretical features. We propose a flexible parameterization approach to include graph-theoretical features as explanatory variables in a prediction model. In particular, flexible functional weight functions of the threshold value determined by statistical goodness-of-fit criteria enable us to incorporate uncertainties of network inference in the model. We perform a simulation study to provide evidence for a proof-of-concept in individual-specific networks of a given size, density and with particular, well-defined network properties. We compare the predictive performance of our approach to a more conventional method of selecting a single sparsification threshold based on AIC. We highlight some challenges that need to be addressed before our approach is ready for routine applications and provide recommendations for our proposed approach in an applied data setting.
