Understanding multimorbidity: insights with graphical models.

Abstract

Background: The use of graphical models in the multimorbidity context is increasing in popularity due to their intuitive visualization of the results. A comprehensive understanding of the model itself is essential for its effective utilization and optimal application. This article is a practical guide on the use of graphical models to better understand multimorbidity. It provides a tutorial with a focus on the interpretation of the model structure and of the parameter values. In this study, we analyze data related to a cohort of 214,401 individuals, who were assisted by the Local Health Unit of the province of Padova (north-eastern Italy), collecting information from hospital discharge forms.

Methods: We explain some fundamental concepts, with special attention to the difference between marginal and conditional associations. We emphasize the importance of considering multimorbidity as a network, where the variables involved are part of an interconnected system of interactions, to correct for spurious effects in the analysis. We show how to analyze the network structure learned from the data by introducing and explaining some centrality measures. Finally, we compare the model obtained by adjusting for population characteristics with the results of a stratified analysis.

Results: Using examples from the estimated model, we demonstrate the key differences between marginal and conditional associations. Specifically, we show that, marginally, all variables appear associated, while this is not the case when considering conditional associations, where many variables appear to be conditionally independent given the others. We present the results from the analysis of centrality indices, revealing that cardiovascular diseases occupy a central position in the network, unlike more peripheral conditions such as sensory organ diseases. Finally, we illustrate the differences between networks estimated in subpopulations, highlighting how disease associations vary across different groups.

Conclusion: Graphical models are a versatile tool for analyzing multimorbidity, offering insights into disease associations while controlling for the effects of other variables. This paper provides an overview of graphical models without focusing on detailed methodology, highlighting their utility in understanding network structures and potential subgroup differences, such as gender-related variations in multimorbidity patterns.