Network visualization to discern patterns of relationships between symptoms in dementia

The multidimensional characterization of complex biomedical systems usually demands a large number of cases in order to obtain reliable inferences. Even so, the number of participants in many studies is relatively small as, for example, in typical clinical trials. Here we suggest an approach based on network visualization, combined with resampling, to discern the patterns of relationships among variables. We illustrate how this can be applied to analyze changes in multiple outcomes in people with dementia. The relationships between several dozens of variables were represented by connectivity graphs, drawn by calculating the relative risk of observing a pair of symptoms in an individual to their co-occurrence by chance only. The statistical significance of the relationships was calculated by generating a bootstrap sample. If the null hypothesis (e.g., the relative risks=1 or equivalently, the pointwise mutual information=0) was rejected, the vertices on the graph representing the variables were connected by an edge. The number of edges (the degree of connectivity) reflects the stage of the cognitive impairment, with worse dementia indicated by lower connectivity. Arranging symptoms consistently allows characteristic profiles to be displayed; this in turn can allow patterns of treatment effects to be discerned, with at-a-glance pattern recognition.