Accurate Prediction of Individual Subject Identity and Task -- but Not ASD Diagnosis -- from Functional Connectomes

Poster Presentation
Saturday, May 12, 2018: 11:30 AM-1:30 PM
Hall Grote Zaal (de Doelen ICC Rotterdam)
L. Byrge and D. P. Kennedy, Psychological and Brain Sciences, Indiana University, Bloomington, IN
Background: There is currently much interest in using functional connectivity (FC) MRI to identify neural biomarkers of ASD. However, for FC-based biomarkers to be useful, we need to know whether FC measurements are reliable in individuals with ASD and whether FC can robustly differentiate diagnostic groups (ASD vs. controls). Progress toward addressing these issues should also provide insights into the biological bases of ASD, including the neural heterogeneity or homogeneity of the condition.

Objectives: To evaluate FC reliability and predict diagnosis group in a sample of adults with ASD and controls using extensively and densely sampled fMRI data from each subject.

Methods: We acquired multiband resting-state and video-watching scans (813ms TR) from 29 adult controls and 22 matched individuals with ASD across multiple scan sessions, comprising over 2 hours of functional data per individual. Preprocessing, denoising, parcellation, and censoring are described in Byrge & Kennedy (in revision; see also Burgess et al., 2016). We computed correlations between each pair of FC matrices to assess similarity of FC. We used general linear mixed models to assess group differences in reliability and in similarity to NT scans while controlling for noise covariates. We used “connectome fingerprinting” (Finn et al., 2015) to predict identity, diagnosis, and scan type (rest/video) from maximal pairwise similarity. We also trained a classifier on the most informative FC edges to predict diagnosis and scan type while also controlling for noise covariates, extending the method of Shen et al. (2017). For visualization, we conducted multi-dimensional scaling.

Results: Within-individual consistency of functional connectomes did not differ across groups (rest/video, p=.14/.33, n.s.), indicating equivalent reliability across scans. Individual connectomes were highly distinct from other individuals, such that we could predict subject identity with 98% accuracy (see also Fig. 1, with subject scans generally clustering together). Using the same fingerprinting method, however, we could not predict subject diagnosis (56.8% accuracy, see also Fig. 2a), even though scans from the ASD group were reduced in similarity to NT scans (rest/video, p=0.026/.017). In contrast, this same method could be used to predict scan type (rest/video) with 97.6% accuracy (see also clear clusters in Fig. 2b), indicating that scan type elicited a stronger common FC pattern than diagnosis. Similar results were obtained using classifiers trained on only the most informative FC edges (60% accuracy for diagnosis; 96.8% accuracy for scan type). MDS results (Fig. 2a) depict scans from individuals with ASD spread farther apart from NT scans (i.e., more dissimilar) in all directions, without forming a common cluster, suggesting heterogeneous presentations of functional connectivity in ASD.

Conclusions: Our findings indicate that, at the spatial and temporal scale analyzed, functional connectomes in individuals with ASD present heterogeneously (consistent with our previous work; Byrge, Dubois, et al., 2015), rather than sharing a common pattern of FC abnormality. In other words, when connectomes from individuals with ASD differ from those of controls, they differ idiosyncratically. This suggests that successful biomarkers will require employing and developing analytic techniques sensitive to idiosyncratic (but stable) neural presentations.