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

Area of Science:

• Artificial Intelligence
• Sleep Medicine
• Biomedical Engineering

Background:

• Accurate sleep stage classification is crucial for diagnosing sleep disorders.
• Existing artificial intelligence (AI) methods often rely on limited data or fewer signals.
• Foundational transformer models offer potential for processing complex, longitudinal biological data.

Purpose of the Study:

• To evaluate a foundational transformer model (PFTSleep) for sleep stage classification using 8-hour, multichannel polysomnogram (PSG) data.
• To compare the performance of PFTSleep against existing AI methods for sleep stage classification.

Main Methods:

• Trained a self-supervised foundational transformer (PFTSleep) on 8-hour, 125 Hz PSG data with 7 signals (brain, movement, cardiac, oxygen, respiratory).
• Used PFTSleep encodings to train a separate sleep stage classification model without altering foundational transformer weights.
• Validated and tested the model on the Sleep Heart Health Study (SHHS) and Multi-Ethnic Study of Atherosclerosis (MESA) datasets.

Main Results:

• PFTSleep achieved high Area Under the Receiver Operating Characteristic Curve (AUROC) scores (0.95-0.99) and Area Under the Precision-Recall Curve (AUPRC) scores (0.40-0.82) across sleep stages on the SHHS validation set.
• On the MESA test set, AUROC scores ranged from 0.77-0.96 and AUPRC scores from 0.16-0.65.
• PFTSleep demonstrated improved macro evaluation scores compared to the longest context window state-of-the-art model, including a 3.7% increase in sensitivity.

Conclusions:

• Foundational transformer models leveraging full-night, multi-channel PSG data significantly improve sleep stage classification.
• PFTSleep's approach offers a more robust and accurate method for AI-driven sleep analysis.
• This advancement has implications for improved diagnosis and management of sleep disorders.