Multivariate biosignals, such as electroencephalography (EEG), polysomnography (PSG), and electrocardiography (ECG), pose challenges in modeling due to their long-range temporal dependencies and complex spatial correlations. Graph neural networks (GNNs) are widely employed for modeling multivariate time series data, with two prominent frameworks: time-and-graph and time-then-graph. In this paper, we propose GraphS4mer, a novel graph neural network architecture tailored for biosignal classification tasks. The model consists of a graph structure learning layer and a Graph Structured State Space Sequence (S4) layer. Moreover, the model adopts a parameterization strategy to enhance the model's versatility and adaptability across diverse datasets and scenarios.
Link to original repository: https://github.com/tsy935/graphs4mer.git