Prediction of Language and Social Communication Deficits from fMRI Functional Connectivity in a Cross-Diagnostic Developmental Sample
Poster D57 in Poster Session D with Social Hour, Friday, October 7, 5:30 - 7:15 pm EDT, Millennium Hall
Also presenting in Poster Slam D, Friday, October 7, 5:15 - 5:30 pm EDT, Regency Ballroom
This poster is part of the Sandbox Series.
Sara Sanchez-Alonso1, Yuqing Cai1,2, Richard Aslin1,2; 1Haskins Laboratories, 2Yale University
Impairments in language and social communication skills are common in early development and form the basis for deficits in social interaction and learning. These deficits are often cross-diagnostic and are present in a range of disorders, such as social communication disorder (SCD) and autism spectrum disorder (ASD). Here, we investigate functional connectivity (FC) differences during naturalistic movie-watching that underlie language and social communication deficits in a large cross-diagnostic developmental sample. We address the following question: To what extent are FC patterns predictive of diagnostic status and language deficits? We analyzed FC fMRI data acquired during movie-watching in a cohort of 368 children and young adults (123 individuals with ASD, 116 individuals with a language disorder and 129 controls) aged 6-20 from the Healthy Brain Network dataset (Alexander et al. 2017). The dataset was preprocessed with the Human Connectome Project minimal preprocessing pipelines (Glasser et al. 2013). To isolate parcel-level and network-level signals we used the whole-brain CAB-NP parcellation (Ji et al. 2019) derived from the HCP atlas (Glasser et al. 2016). The data were analyzed using PrimeNet, a predictive modelling approach that quantifies binary and continuous phenotypic predictions from brain-wide FC patterns (Sanchez-Alonso et al., 2021). Planned analyses include running 1,000 iterations of PrimeNet’s multi-level model to select FC edges (pairwise correlations) for each iteration. The edge-features selected in each iteration will be used to train single-feature SVM binary classifiers using leave-one-out cross-validation. We will apply the trained models to held-out data across the 1,000 iterations. Finally, we will identify the edges across all 1,000 iterations for the main effect of diagnostic status. Preliminary analyses with a subset of the data (30 individuals with a language disorder, 30 individuals with ASD, and 30 age- and IQ-matched controls) using a single iteration of the multi-level model show a classification accuracy of ~70% to distinguish diagnostic status (ASD/language disorder versus controls). In turn, we will derive a dimensionality-reduced symptom space via principal component analysis (PCA) across core language deficit symptoms and social communication skills. Specifically, we will quantify the correlations between 77 symptom measures across all subjects that received a diagnosis (n=239) and will quantify the variance explained by each of the components from a PCA performed using all symptom measures. Using PrimeNet, we will test whether symptom axes map onto distinct FC patterns that allow prediction of language and social communication skills at the individual-subject level. Collectively, there are two key potential outcomes of the planned analyses. First, we predict that there will be a set of FC patterns that can distinguish individuals with a language disorder and ASD (versus controls) in a large-scale developmental sample. These data would provide evidence for a ‘core’ functional network organization in development that varies by diagnostic status. Second, we expect to show that heterogenous language and social communication deficits can be reduced into a low-rank symptom solution that is cross-diagnostic. These derived symptom axes are expected to map onto distinct neural patterns, which are predictive at the individual-subject level.
Topic Areas: Disorders: Developmental, Computational Approaches