ProtSleepNet: An interpretable model for sleep staging based on prototype-learning

This is the authors' abstract. We don't add key points for this paper.

Automatic sleep staging plays a critical role in the diagnosis and treatment of sleep disorders, especially deep learning-based automatic sleep staging methods. However, most existing models provide limited insight into the waveform patterns underlying their decisions, and current interpretability techniques seldom offer explicit, clinically meaningful decision criteria. To address above issues, we propose an interpretable network named ProtSleepNet, which provides both staging result and the corresponding decision. The network derives latent representations during training to determine waveform prototypes that correspond to each sleep stage. We propose a Learnable Gabor-wavelet Dual-stream (LG-WDS) module for feature extraction. We design a Structured Prototype Reasoning (SPR) module. It constructs a comprehensive prototype space by combining Gabor, Fourier and learnable bases. Furthermore, we introduce a Subspace Cross-Matching (SCM) module to align EEG features with the prototypes in mutually latent subspaces based on cross-attention mechanism. The module explicitly quantifies the similarity to derive the final decision. During the training process, a set of loss functions is employed to supervise the diversity and robustness of the waveform prototypes. Meanwhile, the learned waveform prototypes are visualized via occlusion sensitivity analysis. Finally, the proposed ProtSleepNet is evaluated on three public datasets: Sleep-EDF-20, Sleep-EDF-78 and the Sleep Heart Health Study (SHHS), achieving macro F1-scores of 81.2%, 78.6% and 80.5%, and overall accuracies of 86.6%, 84.1%, and 86.3% respectively. This demonstrates that the interpretable method can provide robust staging results.