Multidimensional Neuroanatomical Subtyping of Autism Spectrum Disorder

Abstract

Autism spectrum disorder (ASD) is a group of neurodevelopmental disorders with multiple biological etiologies and highly variable symptoms. Using a novel analytical framework that integrates cortex-wide MRI markers of vertical (i.e., thickness, tissue contrast) and horizontal (i.e., surface area, geodesic distance) cortical organization, we could show that a large multi-centric cohort of individuals with ASD falls into 3 distinctive anatomical subtypes (ASD-I: cortical thickening, increased surface area, tissue blurring; ASD-II: cortical thinning, decreased distance; ASD-III: increased distance). Bootstrap analysis indicated a high consistency of these biotypes across thousands of simulations, while analysis of behavioral phenotypes and resting-state fMRI showed differential symptom load (i.e., Autism Diagnostic Observation Schedule; ADOS) and instrinsic connectivity anomalies in communication and social-cognition networks. Notably, subtyping improved supervised learning approaches predicting ADOS score in single subjects, with significantly increased performance compared to a subtype-blind approach. The existence of different subtypes may reconcile previous results so far not converging on a consistent pattern of anatomical anomalies in autism, and possibly relate the presence of diverging corticogenic and maturational anomalies. The high accuracy for symptom severity prediction indicates benefits of MRI biotyping for personalized diagnostics and may guide the development of targeted therapeutic strategies.

Figure 1.

ASD subtyping. (A) Top: The 4 MRI-based features (cortical thickness, surface area, intensity contrast, geodesic distance) used for subject-wise profiling. Bottom: Three classes (ASD-I, II, III) optimally described our multi-centric cohort. Features are corrected for age and site effects. Shown are differences in MRI features compared to controls (increases/decreases in red/blue). Cortex-wide significances were corrected using the false discovery rate procedure, furthermore adjusted by the number of features (q_FDR < 0.05/4 = 0.0125). (B) Symptom severity profiles of ASD classes based on total and domain-specific ADOS scores (i.e., social interaction, communication, repeated behavior). **Difference between subgroups at q_FDR < 0.05; *Difference at P < 0.05 uncorrected.

Figure 2.

Functional connectivity analyses. (A) We generated functional networks between regions-of-interest (ROIs) using an ad hoc meta-analysis of previous studies on mentalizing (upper panel) and communication (lower panel), based on Neurosynth.org. Networks in healthy controls are presented for reference on the left, with intensity values reflecting inter-regional connectivity strength. (B) Functional connectivity differences between the neuroanatomically defined ASD subtypes (see Fig. 1) and controls, with increases/decreases shown in red/blue. •Findings significant after network-wide FDR correction (q_FDR < 0.05/2 = 0.025), furthermore adjusted by the number of networks.

Figure 3.

Automated prediction of symptom severity. A supervised learning paradigm was used to predict ADOS scores in individual subjects. Shown are observed and predicted ADOS scores for a classifier that was either naïve of the subtyping (A) or informed by it (B). The scatter plots show the mean predicted scores for each ASD individual, averaged across the 100 iterations of the 10-fold cross-validation. The observation-prediction correlation at each iteration is displayed in thin gray, together with a 95% confidence interval. Prediction performance was benchmarked using 3 indices: Pearson correlation [r], statistical significance [p], and mean absolute error [MAE]. Observation-prediction correlations were also presented across subgroups. Finally, selected features (those that were chosen in more than 50% of the 100 cross-validations) were mapped back to cortical surface models (i.e., red: thickness, yellow: intensity contrast, green: surface area, purple: geodesic distance).