RobustSleepNet: Transfer Learning for Automated Sleep Staging at Scale.

Area of Science:

• Neuroscience
• Medical Informatics
• Artificial Intelligence

Background:

• Sleep disorder diagnosis depends on polysomnography (PSG) analysis, with sleep stage classification being a crucial preliminary step.
• Current manual sleep staging is time-consuming and expensive, prompting the development of automated methods.
• Existing automated approaches struggle with diverse PSG montages and demographic variations common in clinical settings.

Purpose of the Study:

• To develop a deep learning model for automatic sleep stage classification that is robust to arbitrary PSG montages and demographic differences.
• To create a model that can be used effectively in clinical sleep labs without extensive customization.

Main Methods:

• Introduction of RobustSleepNet, a deep learning model designed for automatic sleep stage classification.
• Training and evaluation using a leave-one-out approach on 8 heterogeneous sleep staging datasets to ensure robustness.
• Assessment of the model's performance on unseen datasets with different demographics and PSG montages.

Main Results:

• RobustSleepNet achieves 97% of the F1 score of a dataset-specific model when evaluated on unseen data.
• The model demonstrates high performance across arbitrary PSG montages and different demographic groups.
• Fine-tuning RobustSleepNet on a small portion of unseen data further improves performance by 2% compared to training a model from scratch for that dataset.

Conclusions:

• RobustSleepNet offers a viable solution for high-quality, out-of-the-box automatic sleep staging in any clinical environment.
• The model's robustness to montage and demographics significantly addresses limitations of current automated sleep staging methods.
• Fine-tuning provides a pathway to achieve state-of-the-art performance for specific patient populations.