EGCN-TSD: Explainable GCN for Time Series Data and Its Applications to the Study of Brain Development

Functional connectivity is vital in understanding the dynamic process of the brain function network during its development. At present, graph convolutional networks (GCNs) have received extensive attention in the study of brain connectivity during development. Approaches of GCNs utilizing spatio-temporal data generated by medical technologies have been employed for the study of brain development and brain diseases. However, many of these studies disregard both the time-series information and the high-dimension but small-sample-size characteristics of the data, in addition to the explainability of the models used. In this study, we present an explainable GCN for time series data. The regions of interest of human brain are regarded as GCN nodes, and frequency-domain embedding is derived using fast Fourier transform and dictionary learning to effectively extract distinct frequency features and reduce the dimensionality of data. The functional connectivity matrix obtained with truncated nuclear norm regularization is used as the initial value of GCN edges, and attention mechanism is incorporated to capture the dynamic connections, thus to improve the feature extraction and learning abilities of GCN, and reduce the model complexity. For explainability, gradient-weighted class activation mapping and the transition matrix of importance are introduced to identify the key brain regions and connections. Experimental results show that adults’ resting-state networks (RSNs) exhibit denser within-network connectivity, while children’s connections between RSNs are more dispersed. During development, functional partitions of the human brain network will move from segregated to integrated, i.e., gradually become aggregated and modular. The discoveries reflect progressively maturing brain functions, which further uncovers significant changes occur in functional connectivity that are likely related to deduction, reasoning and flexible abilities during development.