For the goal of individualized medicine, it is critical to have clinical phenotypes at hand which represent the individual pathophysiology. However, for most of the utilized phenotypes, two individuals with the same phenotype assignment may differ strongly in their underlying biological traits. In this paper, we propose a definition for individualization and a corresponding statistical operationalization, delivering thereby a statistical framework in which the usefulness of a variable in the meaningful differentiation of individuals with the same phenotype can be assessed. Based on this framework, we develop a statistical workflow to derive individualized phenotypes, demonstrating that under specific statistical constraints the prediction error of prediction scores contains information about hidden biological traits not represented in the modeled phenotype of interest, allowing thereby internal differentiation of individuals with the same assigned phenotypic manifestation. We applied our procedure to data of the population-based Study of Health in Pomerania to construct a refined definition of obesity, demonstrating the utility of the definition in prospective survival analyses. Summarizing, we propose a framework for the individualization of phenotypes aiding personalized medicine by shifting the focus in the assessment of prediction models from the model fit to the informational content of the prediction error.
Keywords: Individualization; directed acyclic graphs; individualized medicine; measurement error; obesity; personalized medicine; prediction modelling.