BrainGET: Decoding Brain Dynamic Functional Connectivity Implicated in ADHD Subtypes

1395 

Abstract Submission 

Deepank Girish¹, Yi Hao Chan¹, Jing Xia¹, Sukrit Gupta², Jagath Rajapakse¹

¹School of Computer Science and Engineering, Nanyang Technological University, Singapore, Singapore, ²Department of Computer Science and Engineering, Indian Institute of Technology, Ropar, India

Deepank Girish  
School of Computer Science and Engineering, Nanyang Technological University
Singapore, Singapore

Yi Hao Chan  
School of Computer Science and Engineering, Nanyang Technological University
Singapore, Singapore

Jing Xia  
School of Computer Science and Engineering, Nanyang Technological University
Singapore, Singapore

Sukrit Gupta  
Department of Computer Science and Engineering, Indian Institute of Technology
Ropar, India

Jagath Rajapakse  
School of Computer Science and Engineering, Nanyang Technological University
Singapore, Singapore

Subtype-specific differences in functional connectivity (FC) have been well-documented for ADHD [2]. Recent studies have revealed strong associations between dynamic FC (dFC) changes and behavioral/cognitive functions [7, 9]. While static FC has been thoroughly investigated in this regard, more needs to be done to study the subtype-specific changes in dFC. We propose BrainGET (Graph neural network Ensembles with Transformers) architecture to capture different extents of spatiotemporal relationships in dFC. BrainGET outperforms state-of-the-art models in ADHD detection and GNNExplainer was used on BrainGET to generate potential subtype-specific biomarkers from the ADHD-200 dataset.

ADHD-200 contains rs-fMRI scans from 279 individuals diagnosed with ADHD and 488 age-matched typical controls, collected from 4 sites: NI, NYU, OHSU and PKU. Also, there are 3 subtypes of ADHD: Hyperactive-Impulsive, Inattentive and Combined (Figure 1c). dFC matrices were generated using a sliding window approach and used as inputs to our proposed model. BrainGET integrates Graph Convolutional Network (GCN) and Graph Isomorphism Network (GIN) in an ensemble framework, then uses Transformers to capture temporal dynamic patterns (Figure 1a). GCN and GIN differ in how node information is spatially aggregated, while Transformers provide an effective way of capturing temporal dynamics from FC. Using an ensemble allows the model to learn the optimal way of combining these spatiotemporal representations.

We compared BrainGET's site-specific classification performance (healthy vs ADHD) to current state-of-the-art dFC models based on GNNs and Transformer architectures, including: DGCN [8], ST-GCN [3] and STAGIN [4] (Figure 1b). Our model was statistically significant (with paired t-test p-value < 0.05) and outperformed existing methods due to the combined expressive power of GCN and GIN.
For model explainability, attention scores were used to find salient sliding windows from both GCN and GIN component of BrainGET, for each subtype across all 4 sites. We subsequently applied GNNExplainer to generate saliency scores which reveal the contribution of each node to the classification task. Saliency scores derived from the GCN and GIN components were combined using the same weights as utilized in the ensemble of classification predictions.
Decoding was done for each site, but analysis and figures exclude site NI due to small sample size. We found that the connection between insula and extra-nuclear is most salient for both sites NYU and OHSU (Figure 2). Notably, it is not present in PKU, possibly due to different races (Western vs Chinese). Additionally, its presence in both combined and inattentive subtypes and absence in hyperactive subtype suggests that it could be an FC feature specific to the inattentive subtype. While the combined and inattentive subtypes have distinct salient connections, the involvement of insula and extra-nuclear regions, known for their role in consciousness and cognitive functions within the brain, aligns with findings from prior ADHD studies [1, 5].
When saliency scores are consolidated at the level of brain modules, we find significant similarities between the combined and inattentive subtypes (with all p-values < 0.05): Person's r = 0.97 for NYU, r = 0.84 for OHSU and r = 0.71 for PKU. Correlations between combined and hyperactive subtypes were moderate (r = 0.45 for NYU, r = 0.73 for OHSU), correlation between inattentive and hyperactive subtypes were even lower.

Put together, our results from decoding BrainGET suggest that the combined and inattentive subtypes share significant similarities in dFC (in particular, connections between insular and extra-nuclear regions), both at the level of ROIs and brain modules. This could suggest that combined subtypes are dominated by inattentive traits, or that dFC does not differentiate these clinically derived subtypes.

Classification and Predictive Modeling ¹

Connectivity (eg. functional, effective, structural)

fMRI Connectivity and Network Modeling ²

BOLD fMRI

Attention Deficit Disorder

FUNCTIONAL MRI

Machine Learning

Other - Dynamic Functional Connectivity; Subtyping; Modules