Autism Subgroups Defined By Heterogeneity in Performance on an Advanced Mentalizing Test
Objectives: Here we use unsupervised hierarchical clustering to stratify adults on the autism spectrum into discrete subgroups, based on their performance on an advanced mentalizing test.
Methods: Two datasets (discovery and replication) were used to test whether discrete autism mentalizing subgroups emerge in a replicable fashion. The discovery dataset consisted of 190 adults with autism (96 males, 94 females) and 251 typically-developing controls (TD; 130 males, 121 females) who completed the Reading the Mind in the Eyes test (RMET) online as part of the Cambridge Autism Research Database (CARD). The replication dataset consisted of participants from the MRC AIMS Consortium dataset (n=123 autism; 85 male, 38 female; n=128 TD; 87 male, 41 female). For both datasets, we used unsupervised two-way hierarchical clustering to cluster the data along both dimensions of RMET items and autism subjects. We tested the hypothesis that similar discrete subgroups would emerge across both discovery and replication datasets. We then compared autism subgroups to TD using RMET total scores and estimate replication Bayes Factors (repBF) to quantify evidence for replicability (repBF ~ 1 indicates no evidence supporting replication; repBF>100 indicates extremely strong evidence supporting replication).
Results: Across both discovery and replication datasets there was evidence for 2 discrete autism subgroups that can generally be characterized as ‘Impaired’ or ‘Good’. The ‘Good’ subgroup showed no consistent difference in RMET total scores compared to the TD group (Discovery: t = 1.82, p = 0.06, Cohen’s d = 0.21; Replication: t = -3.72, p = 0.0002, Cohen’s d = 0.52; repBF = 0.22). However, the ‘Impaired’ subgroup had RMET total scores that were massively reduced in a replicable manner compared to the TD group (Discovery: t = -16.96, p = 4.10e-47, Cohen’s d = 2.10; Replication: t = -13.04, p = 3.13e-27, Cohen’s d = 2.36; repBF = 1.63e+25).
Conclusions: The RMET can be a valuable tool for parsing heterogeneity in mentalizing in adults with autism. After stratification, there is a useful distinction between autism individuals with intact versus impaired mentalizing ability. Stratification by these mentalizing subgroups may be useful in future work with clinical and translation aims. For example, in direct extensions of this work, we will directly test these subgroups for differentiation within mentalizing circuitry with existing task and resting state fMRI data from the MRC AIMS dataset.