Graph-Based Learning of Coupled Neurovascular–Biochemical Dynamics from High-Dimensional Signals

Abstract

The interaction of such components of neural activity, vascular responses and biochemical processes is important in the interpretation of complex physiological dynamics. Nevertheless, the current ways of analysis tend to address these modalities separately, or through the use of clustered associations rather than dynamic, inter-domain relationships of high-dimensional data. In this paper, we provide a graph-based learning model that learning in both coupled neurovascular and biochemical dynamics with high-dimensional time series. The suggested method implies multimodal physiological units to be presented as nodes in a latent interaction graph, and each of the nodes is connected by edge weights that capture learned functional and biochemical interactions. The framework can concurrently learn interaction structure and temporal dynamics through the combination of message passing and graph learning in an end-to-end manner. To encourage interpretability and achieve good predictions, a sparsity-regularized objective is defined. Through large-scale experiments with multimodal data at a large scale, it has been proved that the method in question outperforms the current state of the art on accuracy and stability of predictions as well as displaying high-level patterns of cross-domain couplings. These findings demonstrate that graph-based intelligent modelling is a scalable and interpretable way of analysing complex coupled physiological systems.