U-Sleep's resilience to AASM guidelines.

Area of Science:

• Sleep medicine
• Artificial intelligence in healthcare
• Biomedical signal processing

Background:

• The American Academy of Sleep Medicine (AASM) provides guidelines for standardizing sleep scoring procedures.
• Automated sleep scoring systems traditionally rely heavily on these AASM standards.
• Deep learning methods have shown superior performance over classical machine learning for sleep scoring.

Purpose of the Study:

• To investigate the performance of a deep learning sleep scoring algorithm (U-Sleep) with non-conventional data inputs.
• To determine if strict adherence to AASM guidelines is necessary for high-performance automated sleep scoring.
• To evaluate the impact of multi-center data versus single-center data on model performance.

Main Methods:

• Utilized U-Sleep, a state-of-the-art deep learning algorithm for sleep scoring.
• Experimented with clinically non-recommended EEG derivations and without chronological age information.
• Trained and validated models using 28,528 polysomnography studies from 13 diverse clinical studies.
• Compared performance of models trained on single large cohorts versus multi-center data.

Main Results:

• U-Sleep achieved strong performance even with non-conventional EEG derivations and without age data.
• The necessity of strictly adhering to AASM guidelines for optimal performance was challenged.
• Training on multi-center data consistently yielded better performing models compared to single-center training, even with large, heterogeneous single cohorts.

Conclusions:

• Deep learning models for sleep scoring can be robust and perform well without full reliance on traditional AASM guidelines or all recommended data.
• Multi-center data remains crucial for enhancing the generalizability and performance of automated sleep scoring systems.
• The findings suggest flexibility in data utilization for developing effective AI-driven sleep analysis tools.