MuRAt-CAP-Net: A novel multi-input residual attention network for automated detection of A-phases and subtypes in cyclic alternating patterns

Abstract

Cyclic Alternating Pattern (CAP) is an essential biomarker for evaluating sleep microstructure, analyzing sleep stability and detecting various sleep disorders. The manual scoring of the CAP A-phase and its subtypes (A1, A2, A3) is a time-consuming, complex and expert-dependent. In this study, we propose a novel deep learning model named the multi-input residual attention CAP network (MuRAt-CAP-Net) for the automated detection of CAP A-phase and its subtypes. MuRAt-CAP-Net, with its multi-input architecture, simultaneously processes signals from four EEG channels (C4-P4, F4-C4, Fp2-F4, P4-O2) originating from different cortical areas. Additionally, the integrated attention mechanisms enable the model to focus on the most relevant features. The performance of MuRAt-CAP-Net was evaluated on both balanced and imbalanced datasets using a 5-fold cross-validation strategy. For A-phase classification, the model achieved an accuracy of 81.26 % and an F1-score of 81.13 % on the balanced dataset, while achieving 83.68 % accuracy and 88.38 % F1-score on the imbalanced dataset. For subtype classification, the model achieved an accuracy of 83.34 % and an F1-score of 83.38 % on the balanced dataset, and 87.31 % accuracy and 84.64 % F1-score on the imbalanced dataset. Compared to state-of-the-art methods, MuRAt-CAP-Net demonstrated superior performance in the detection of CAP A-phases and their subtypes. Furthermore, to enhance interpretability, Grad-CAM was applied to visualize the temporal and spectral focus of the MuRAt-CAP-Net’s decisions, revealing physiologically consistent patterns and supporting the clinical reliability of the model. Additionally, this study provides a comprehensive analysis of the impact of different EEG channel combinations, input window durations, and attention mechanisms on model performance.