WaveSleepNet: An Interpretable Network for Expert-Like Sleep Staging.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Medical Informatics

Background:

• Deep learning models excel at automatic sleep staging but lack transparency, hindering clinical use.
• Interpretable AI is crucial for medical applications, especially in sleep medicine where understanding reasoning is vital.

Purpose of the Study:

• To develop an interpretable neural network, WaveSleepNet, for automatic sleep staging.
• To enable sleep staging decisions that align with clinical expert reasoning and guidelines.

Main Methods:

• WaveSleepNet utilizes latent space representations to identify characteristic wave prototypes for different sleep stages.
• A scoring system quantifies the proximity of signal patches to learned prototypes for decision-making.
• Ensemble loss functions and occlusion sensitivity analysis enhance prototype robustness and interpretability.

Main Results:

• WaveSleepNet achieves state-of-the-art performance in sleep staging across three public datasets.
• Case studies confirm WaveSleepNet's decision-making aligns with American Academy of Sleep Medicine (AASM) guidelines.
• The model provides systematic explanations for misidentified sleep stages, aiding in error analysis.

Conclusions:

• WaveSleepNet offers a transparent and interpretable alternative to black-box deep learning models for sleep staging.
• The model's interpretability facilitates clinical validation, adoption, and refinement by sleep specialists.
• This approach enhances trust and understanding in AI-driven sleep analysis tools.