Organizing heterogeneous samples using community detection of GIMME-derived resting state functional networks

Clinical investigations of many neuropsychiatric disorders rely on the assumption that diagnostic categories and typical control samples each have within-group homogeneity. However, research using human neuroimaging has revealed that much heterogeneity exists across individuals in both clinical and control samples. This reality necessitates that researchers identify and organize the potentially varied patterns of brain physiology. We introduce an analytical approach for arriving at subgroups of individuals based entirely on their brain physiology. The method begins with Group Iterative Multiple Model Estimation (GIMME) to assess individual directed functional connectivity maps. GIMME is one of the only methods to date that can recover both the direction and presence of directed functional connectivity maps in heterogeneous data, making it an ideal place to start since it addresses the problem of heterogeneity. Individuals are then grouped based on similarities in their connectivity patterns using a modularity approach for community detection. Monte Carlo simulations demonstrate that using GIMME in combination with the modularity algorithm works exceptionally well--on average over 97% of simulated individuals are placed in the accurate subgroup with no prior information on functional architecture or group identity. Having demonstrated reliability, we examine resting-state data of fronto-parietal regions drawn from a sample (N = 80) of typically developing and attention-deficit/hyperactivity disorder (ADHD) -diagnosed children. Here, we find 5 subgroups. Two subgroups were predominantly comprised of ADHD, suggesting that more than one biological marker exists that can be used to identify children with ADHD based from their brain physiology. Empirical evidence presented here supports notions that heterogeneity exists in brain physiology within ADHD and control samples. This type of information gained from the approach presented here can assist in better characterizing patients in terms of outcomes, optimal treatment strategies, potential gene-environment interactions, and the use of biological phenomenon to assist with mental health.