PIANet: a parallel interactive attention network for multi-channel PSG-based sleep staging

Background.Sleep is crucial to human health and quality of life. Existing methods often struggle to simultaneously capture cross-channel correlations and spatial dependencies in multi-channel polysomnography (PSG) signals, which may limit feature fusion.Objective.This study aims to enhance multi-channel PSG-based sleep staging by explicitly modeling the interaction between channel and spatial feature importance within a unified framework.Methods.We propose a parallel interactive attention network (PIANet). This network jointly characterizes the importance of features in both channel and spatial dimensions while enabling interaction-based fusion of channel and spatial features. Specifically, the model uses bidirectional gated recurrent units to extract epoch-level features and sequence-level features, uses a parallel interactive attention (PIA) mechanism to represent channel-wise and spatial importance, and further integrates multi-channel features through a convolutional fusion module (ConvBlock).Results.Compared with baseline models, PIANet achieves competitive classification performance in three public datasets, supporting its effectiveness.Conclusion.PIANet provides an interaction-aware framework for multi-channel PSG-based sleep staging by jointly modeling channel and spatial attention. At its current stage, the method is primarily intended for computer-assisted scoring in laboratory PSG settings. Future research directions will include robustness to signal degradation, adaptation to heterogeneous recording conditions, and lightweight online deployment.