Different functional neural substrates for good and poor language outcome in autism

Abstract

Autism (ASD) is vastly heterogeneous, particularly in early language development. While ASD language trajectories in the first years of life are highly unstable, by early childhood these trajectories stabilize and are predictive of longer-term outcome. Early neural substrates that predict/precede such outcomes are largely unknown, but could have considerable translational and clinical impact. Pre-diagnosis fMRI response to speech in ASD toddlers with relatively good language outcome was highly similar to non-ASD comparison groups and robustly recruited language-sensitive superior temporal cortices. In contrast, language-sensitive superior temporal cortices were hypoactive in ASD toddlers with poor language outcome. Brain-behavioral relationships were atypically reversed in ASD, and a multimodal combination of pre-diagnostic clinical behavioral measures and speech-related fMRI response showed the most promise as an ASD prognosis classifier. Thus, before ASD diagnoses and outcome become clinically clear, distinct functional neuroimaging phenotypes are already present that can shed insight on an ASD toddler's later outcome. VIDEO ABSTRACT.

Fig. 1. Developmental trajectories for language, non-verbal cognitive, or autism symptom measures

This figure shows developmental trajectories for all groups (TD, red; LD/DD, green; ASD Good, blue; ASD Poor, purple) on the Mullen expressive language (EL) (A), receptive language (RL) (B), visual reception (VR) (C) subscales, and the Autism Diagnostic Observation Schedule (ADOS). Plots show the group-level trajectory (solid line) estimated from mixed-effect modeling after taking into account individual-level trajectories (dotted lines, unfilled circles). The colored bands indicate the 95% confidence band around the group-level trajectory.

Fig. 2. Functional neural abnormalities in autism with poor language outcome and development

Panels A–D show the difference in BOLD percent signal change for the contrast of all speech conditions versus rest for meta-analytically defined canonical neural systems for language. Errorbars represent 95% confidence intervals. Panel E (top row) shows the full spatial extent of the NeuroSynth ‘language’ meta-analysis map along with activation observed at a whole-brain corrected level (FDR q<0.05) within each group individually. The second to last row of panel E shows the whole-brain analysis for the specific contrast of All Other Groups > ASD Poor. The last row of panel E shows the conjunction overlap between the NeuroSynth ‘language’ map and the whole-brain contrast of All Other Groups > ASD Poor. In all panels the coloring across groups is red = TD, green = LD/DD, blue = ASD Good, and purple = ASD Poor.

Fig. 3. Multi-Voxel Pattern Similarity Analyses

Panels A and B show pattern similarity analyses comparing each group’s second-level group activation maps (i.e. t-statistic maps) with all other groups. These analyses were conducted only on voxels defined based on the NeuroSynth language map (i.e. voxels that passed FDR q<0.01). Panel A shows the similarity matrix (i.e. Pearson correlations) comparing each group’s second-level activation map with every other group. Higher values closer to 1 indicate more pattern similarity, while values closer to 0 indicate no pattern similarity across voxels. The correlation matrix in panel A was converted into a dissimilarity matrix (i.e. 1-r) and entered into canonical multidimensional scaling to reduce the matrix to 2-dimensions (i.e. MDS1 and MDS2) for the purposes of visualization of the separation between ASD Poor and all other groups (panel B). Panels C and D show results of analyses comparing individual subject activation maps (i.e. t-statistic maps) to specific NeuroSynth maps for either ‘Language AND Speech’ or ‘Auditory AND NOT (Language OR Speech)’. Errorbars represent 95% confidence intervals. In all panels the coloring across groups is red = TD, green = LD/DD, blue = ASD Good, and purple = ASD Poor.

Fig. 4. Multivariate brain-behavioral relationship analyses

This figure shows large-scale neural systems where all speech vs rest fMRI response covaries with multivariate behavioral patterns of variation in language ability measured at intake and outcome time-points. Partial least squares correlation analysis highlighted one significant brain-behavioral latent variable pair (LV1; d = 73.90, p = 0.031; 29.60% covariance explained) and the strength of brain-behavioral relationships for LV1 are shown in panels A–B, while the voxels that most robustly express such relationships are shown in panels C–D. Panel A depicts the directionality of brain-behavioral relationships for hot colored brain regions in panels C–D, while panel B depicts the directionality of brain-behavioral correlations for cool colored brain regions in panels C–D. Within the bar plots in panels A–B, bars are stratified by language measure (Mullen Expressive (EL) or Receptive (RL) language or Vineland Communication), and by the developmental time-point at which they were measured (e.g., intake or outcome assessment). Error bars indicate the 95% confidence intervals estimated from bootstrapping (10,000 resamples). The coloring of the bars indicate different groups (red = TD; green = LD/DD; blue = ASD Good; purple = ASD Poor). Stars above specific bars indicate where the brain-behavioral relationship is non-zero (i.e. 95% CIs do not encompass 0) and these are specific relationships that reliably contribute to the overall PLS relationship for LV1. The coloring in panels C–D reflect the bootstrap ratio (BSR) which is analogous to a pseudo z-statistic and can be interpreted accordingly. Only voxels in panels C–D with BSR values greater than 1.96 or less than −1.96 are shown, as these are the primary voxels showing the biggest contributions to the overall pattern being expressed by LV1. Abbreviations: BSR = bootstrap ratio; EL, Mullen Expressive Language; RL Mullen Receptive Language; ASD, autism spectrum disorder; TD, typical development; LD/DD, language/developmental delay.

Fig. 5. Prognostic classifier performance within ASD for models using clinical behavioral data, fMRI data or both at pre-diagnostic ages

This figure shows the performance of 4 classifier models (i.e. partial least squares linear discriminant analyses) using different features for the purpose of distinguishing the ASD Poor from the ASD Good subgroup. All models using behavioral data (e.g., ADOS, Mullen, Vineland) utilize such information from the earliest clinical intake time-point. Information input into the fMRI classifier are percent signal change estimates from the all speech vs rest from all voxels within the NeuroSynth left hemisphere superior temporal cortex ROI. Abbreviations: Acc = Accuracy; Sens, Sensitivity; Spec, Specificity; AUC, area under the curve; ADOS, Autism Diagnostic Observation Schedule.