A foundational transformer leveraging full night, multichannel sleep study data accurately classifies sleep stages

Area of Science:

• Artificial Intelligence in Medicine
• Sleep Science and Technology

Background:

• Accurate sleep stage classification is crucial for diagnosing sleep disorders.
• Existing AI models often require extensive task-specific training data.
• Foundational models offer a promising avenue for generalizable representations in medical data.

Purpose of the Study:

• To evaluate a foundational transformer model for encoding polysomnography (PSG) data.
• To assess the model's ability to perform state-of-the-art sleep stage classification.
• To determine the generalizability of the model across diverse patient cohorts.

Main Methods:

• Developed PFTSleep, a self-supervised foundational transformer utilizing multichannel PSG data (brain, movement, cardiac, oxygen, respiratory).
• Trained on 13,888 sleep studies from multiple cohorts (Sleep Heart Health Study, Wisconsin Sleep Cohort, MrOS Visit 1).
• Validated on 4,169 independent studies (MESA, APPLES, MrOS Visit 2), comparing performance with existing AI methods.

Main Results:

• PFTSleep achieved high Cohen's Kappa scores on independent test sets (0.59-0.75).
• The model demonstrated robust performance across diverse datasets, outperforming other state-of-the-art AI models.
• Task-agnostic PSG representations from the foundational transformer enabled effective sleep stage classification.

Conclusions:

• Full-night, multichannel PSG data encoded by a foundational transformer enables accurate sleep stage classification.
• The PFTSleep model achieves performance comparable to state-of-the-art AI methods.
• This approach shows promise for generalizable and high-performance sleep analysis across varied datasets.