Data-driven identification of subtypes of executive function across typical development, attention deficit hyperactivity disorder, and autism spectrum disorders

Abstract

Background: Impairment of executive function (EF), the goal-directed regulation of thoughts, actions, and emotions, drives negative outcomes and is common across neurodevelopmental disorders including attention deficit hyperactivity disorder (ADHD) and autism spectrum disorder (ASD). A primary challenge to its amelioration is heterogeneity in symptom expression within and across disorders. Parsing this heterogeneity is necessary to attain diagnostic precision, a goal of the NIMH Research Domain Criteria Initiative. We aimed to identify transdiagnostic subtypes of EF that span the normal to impaired spectrum and establish their predictive and neurobiological validity.

Methods: Community detection was applied to clinical parent-report measures in 8-14-year-old children with and without ADHD and ASD from two independent cohorts (discovery N = 320; replication N = 692) to identify subgroups with distinct behavioral profiles. Support vector machine (SVM) classification was used to predict subgroup membership of unseen cases. Preliminary neurobiological validation was obtained with existing functional magnetic resonance imaging (fMRI) data on a subsample (N = 84) by testing hypotheses about sensitivity of EF subgroups versus DSM categories.

Results: We observed three transdiagnostic EF subtypes characterized by behavioral profiles that were defined by relative weakness in: (a) flexibility and emotion regulation; (b) inhibition; and (c) working memory, organization, and planning. The same tripartite structure was also present in the typically developing children. SVM trained on the discovery sample and tested on the replication sample classified subgroup membership with 77.0% accuracy. Split-half SVM classification on the combined sample (N = 1,012) yielded 88.9% accuracy (this SVM is available for public use). As hypothesized, frontal-parietal engagement was better distinguished by EF subtype than DSM diagnosis and the subgroup characterized with inflexibility failed to modulate right IPL activation in response to increased executive demands.

Conclusions: The observed transdiagnostic subtypes refine current diagnostic nosology and augment clinical decision-making for personalizing treatment of executive dysfunction in children.

Keywords: Attention deficit hyperactivity disorder; autism spectrum disorders; functional MRI (fMRI); individual differences; machine learning.

Figure1.

Results of community detection for each sample. For each sample, the top panel shows a spring-embedded graphical representation that is typically used to visualize the network structure of detected communities. Nodes correspond to individuals in each EF subgroup and length of connecting lines to the correlation between individuals. The graph is thresholded for visualization purposes at r=.6, which is sufficient to see clearly the network structure within each subgroup. However, reported results were generated with an unthresholded modularity algorithm as described in Methods and Appendix S1. Each community is color-coded corresponding to the legend in the graphs in the bottom panel, which plot mean Z scores (± SEM) showing the profile of each EF subtype - left to right, ADHD Rating Scale Hyperactivity/Impulsivity and Inattention; BRIEF Inhibit, Shift, Emotional Control, Initiate, Working Memory, Planning and Organization, Organization of Materials, and Monitor, and CBCL Internalizing and Externalizing symptoms. On the Y-axis, Z=0 indicates the sample mean, and higher z scores indicate higher impairment.

Figure2.

Distribution of values of two metrics describing the Combined sample, which show the dimensional spectrum and transdiagnostic composition of each EF subgroup: (A) Area Under the Curve (AUC) across the 12 T-scores is an index of severity of the weaknesses defining the EF subtype, as higher peaks in the profile provide higher AUC values. For each subtype, lower end of the distribution are occupied by TD participants as expected, with ASD and ADHD participants equally represented by ASD and ADHD participants in the remaining distribution. (B) Within-module z-scores (a community version of degree centrality) which indicates each participant’s position within their subgroup, with the positive end of the y axis indicating central positions (individuals who have higher number of group members correlated with them) and the negative end indicating peripheral positions (individuals who have fewer group members correlated with them). TD, ASD, and ADHD participants represent the full range of the distribution.

Figure3.

Results of community detection for the combined TD sample, with top panel showing a spring-embedded graph of the network structure of detected communities (EF subtype), with nodes corresponding to individuals and length of lines to the correlation between individuals in that community. Similar to Figure 1, the graph is thresholded at r=.6 for visualization purposes. Each community is color-coded corresponding to the legend in the graphs in the bottom panel, which plot the profile of mean Z scores (± SEM) for each EF subtype. Domain scores labeled as in Figure 1.

Figure4.

Group average of activation for Salient vs. Familiar distractor trials during the “bottom-up” run of the sustained attention task (A) and activation in right inferior parietal lobule (RIPL) and right insula/inferior frontal gyrus (insula/IFG) by run. FLEX-EMOT subgroup failed to increase RIPL activation in “top-down” run relative to “bottom-up” run (top right).