Background: Older theoretical models of ADHD have largely assumed homogeneity in cognitive functioning. More recent work exemplifies the heterogeneity in cognitive performance in ADHD populations, but there is not much consensus between studies. Additionally, when comparing between ADHD and a typically developing (TD) group, there is an underlying assumption of homogeneity in the TD population.
To explore this heterogeneity, this study used graph theory to segregate TD and ADHD children into subtypes with a variety of neuropsychological measures. Then tested
Methods: (1) community detection to group into clusters/subtypes (2) SVM-based MVPA to test how well diagnostic status of individuals can be identified within the community clusters/subtypes Asking:
- Is there enough information in these neuropsych subgroups to predict cluster membership?
- Does cluster membership allow for more accurate prediction of ADHD diagnostic status than just neuropsychological performance alone?
498 children (TD = 213, ADHD = 285) & 20 neuropsych measures
Results: Used confirmatory factor analysis to make sure that the measures were not excessively redundant settled on 7 latent variables with 1-4 measures each because it most closely aligned with their prior conceptualization of what each task measured and it had an equivalent fit to the five and six factor models ((latent variables: inhibition, working memory, arousal/activation, response variability, temporal information processing, memory span, processing speed))
Can we distinguish between ADHD and TD populations on the basis of neuropsychological task performance? Despite significant differences between ADHD and TD groups in all neuropsych measures, the SVM classifier was not great at sorting individuals by neuropsych performance into ADHD and TD groups (65% accuracy).
Community Detection Community detection on TD population identified four unique communities/subgroups (1) more response variability, (2) reduced working memory/inhibition/span/speed, (3) inaccurate temporal information processing, (4) weak signal detection/altered arousal **TD populations have heterogeneity in neuropsych profiles
Community detection on ADHD population identified six unique communities/subgroups that they overlayed against the TD communities. « see figure3 »
TD and ADHD communities largely not defined by severity of ADHD symptom burden (using Conners-3), IQ, age, sex.
THERE SHOULD BE AN IMAGE HERE BUT IMAGES ARE HARD
In Figure 3, greater z-score is poorer performance. While there are significant differences across all domains (Figure 3A) withpoorer performance in the ADHD group, the variability in individual abilities is really washed out compared to the profiles in Figure 3B.
SVM-based MVPA Used SVM on these community clusters/subgroups. Split ADHD and TD groups in half, replicated the community clusters with these smaller groups, trained SVM on clusters for TD (4) and ADHD (6). SVM was 78% accurate for TD group (subgroup 1, 93%; subgroup 2, 71%; subgroup 3, 71%; subgroup 4, 81.25%) and 77% accurate for ADHD group (subgroup 1, 78%; subgroup 2A, 75%; subgroup 2B, 66.25%; subgroup 3, 77%; subgroup 4A, 88%; subgroup 4B was too small to test) The authors took this to represent the robustness of the community clustering.
Question: What can differences in the ability of the SVM-based MVPA to accurately assign to different subgroups tell us about differences between subgroups?
Taking the subgroups into consideration, can we better distinguish between ADHD and TD populations on the basis of neuropsychological task performance? Each subgroup has a different pattern of atypicality between diagnostic groups.
- Differences between ADHD and TD in response variability and temporal information priming were each found in four groups.
- Differences between ADHD and TD in inhibitory control, speed, and working memory were each found in three groups.
- Differences between ADHD and TD in arousal and span were each identified in two groups.
ADHD classification improved/remained the same for all groups except 4B (compared to the initial 65% accuracy in classifying ADHD vs TD from the neuropsychological profiles without regard for subtypes). group 1: 68% accuracy in classification of ADHD vs TD youth group 2A: 68.5% group 2B: 64.7% group 3: 73.6% group 4A: 84.1% group 4B: 61.9%
Big Ideas:
(1) Underlying heterogeneity in the ‘typically developing’ population exists & heterogeneity in the ADHD population largely nests within this normal variation. (2) Perhaps understanding that atypical patterns are specific to the community clusters/subgroups explains discrepancies in ADHD neuropsych abilities in the literature. For example, in this study inhibition was only impaired in 3/6 ADHD subgroups and working memory was only impaired in 4/6 ADHD subgroups. (3) Diagnostic status can be more easily identified with the SVM-based MVPA classification once the community clusters/subgroups are taken into consideration, which supports the robustness of these community clusters/subtypes.
Considerations:
- The authors specifically mention a need to evaluate the stability in these neuropsych subtypes over time. Perhaps, the EF Study provides a good opportunity for this – with the eventual collection of four years of data.
- The authors also suggest using neuroimaging to inform the clusters and suspect that each group identified here has distinct neurobiological profiles. The EF Study has neuroimaging task data across three EF domains. Perhaps we could pull measures from there - what would contribute most? [[Sidebar: It would be really interesting to replicate Tehila’s EF x ADHD paper (distinct relations between EF domains and ADHD symptom burden) with regard to cognitive subtypes. If only we had the power…]]