Laying the Foundation: Modern Transformers for Gold-Standard Sleep Analysis and Beyond.

Area of Science:

• Artificial Intelligence
• Sleep Medicine
• Machine Learning

Background:

• Accurate sleep assessment is crucial for sleep medicine and research.
• Advancements in AI, particularly transformer architectures, offer new avenues for sleep data analysis.
• Transformers can facilitate data-driven discovery through unsupervised or self-supervised learning.

Purpose of the Study:

• To evaluate optimized transformer architectures for sleep stage classification.
• To establish foundational baselines for transformer models in sleep research.
• To explore the potential of transformers for future data-driven discovery in sleep.

Main Methods:

• Adapted transformer architectures, successful in audio speech recognition, for sleep stage classification.
• Trained and evaluated transformer models on a benchmark 5-class sleep stage classification task.
• Focused on small transformer models designed for self-supervised learning.

Main Results:

• Optimized small transformer models achieved performance comparable to state-of-the-art automated sleep scoring techniques.
• Demonstrated the feasibility of using transformers for supervised sleep stage classification.
• Established a basis for future self-supervised learning applications in sleep.

Conclusions:

• Transformer models optimized for self-supervised learning can effectively perform supervised sleep stage classification.
• These models provide a strong foundation for leveraging large sleep datasets for data-driven discovery.
• The study highlights the potential of AI to advance sleep medicine and research.