Heat Kernels with Functional Connectomes Reveal Atypical Energy Transport in Peripheral Subnetworks in Autism

Abstract

Autism is increasing in prevalence and is a neurodevelopmental disorder characterised by impairments in communication skills and social behaviour. Connectomes enable a systems-level representation of the brain with recent interests in understanding the distributed nature of higher order cognitive function using modules or subnetworks. By dividing the connectome according to a central component of the brain critical for its function (it's hub), we investigate network organisation in autism from hub through to peripheral subnetworks. We complement this analysis by extracting features of energy transport computed from heat kernels fitted with increasing time steps. This heat kernel framework is advantageous as it can capture the energy transported in all direct and indirect pathways between pair-wise regions over 'time', with features that have correspondence to small-world properties. We apply our framework to resting-state functional MRI connectomes from a large, publically available autism dataset, ABIDE. We show that energy propagating through the brain over time are different between subnetworks, and that heat kernel features significantly differ between autism and controls. Furthermore, the hub was functionally preserved and similar to controls, however, increasing statistical significance between groups was found in increasingly peripheral subnetworks. Our results support the increasing opinion of non-hub regions playing an important role in functional organisation. This work shows that analysing autism by subnetworks with the heat kernel reflects the atypical activations in peripheral regions as alterations in energy dispersion and may provide useful features towards understanding the distributed impact of this disorder on the functional connectome.

Neuropsychiatric disease classification using functional connectomics - results of the connectomics in neuroimaging transfer learning challenge

“…The graph Laplacian and its eigenspectrum enable a mapping of discrete data (a network) into vector spaces and manifolds, and their advantages have been greatly investigated, e.g., in fields such as clustering (Chung (1997)). This form of network representation has also found its way into connectomes and may have potential to capture further discriminative information (Abdelnour et al (2014); Chung et al (2016b,a); Schirmer and Chung (2019)). These data filtering or manipulation techniques are effective for their robustness to overfitting (Du et al (2018)) and may be a reason for S5 and S2 achieving the best generalization performance from Task I to Task II.…”

Neuropsychiatric Disease Classification Using Functional Connectomics -- Results of the Connectomics in NeuroImaging Transfer Learning Challenge

“…The graph Laplacian and its eigenspectrum enable a mapping of discrete data (a network) into vector spaces and manifolds, and their advantages have been greatly investigated, e.g., in fields such as clustering [63]. This form of network representation has also found its way into connectomes and may have potential to capture further discriminative information [64,65,66,67]. These data filtering or manipulation techniques are effective for their robustness to overfitting [68] and may be a reason for S5 and S2 achieving the best generalization performance from Task I to Task II.…”

Network Diffusion Embedding Reveals Transdiagnostic Subnetwork Disruption and Potential Treatment Targets in Internalizing Psychopathologies

“…A closely related methodology to that of studying structural eigenmodes is using the graph Laplacian-derived heat kernel of brain networks, a matrix which encodes how much information is transferred between every pair of network nodes after a given time, or diffusion depth. Heat kernel methods have been used to characterize perturbations in brain network information transfer in autism (Schirmer and Chung, 2019) and to predict future adverse motor function resulting from premature birth using white matter structural connectomes (Chung et al, 2016).…”