Predicting Age with Deep Neural Networks from Polysomnograms.

Area of Science:

• Artificial Intelligence
• Sleep Medicine
• Biotechnology

Background:

• Polysomnography (PSG) is crucial for sleep analysis but often relies on time-consuming clinical scoring.
• Existing deep learning models for sleep analysis may be limited by clinical scoring guidelines and bias.
• Developing automated, end-to-end processing frameworks for PSG data is essential for efficient and objective sleep analysis.

Purpose of the Study:

• To design a novel deep learning framework for end-to-end polysomnogram processing.
• To enable analysis of whole-night sleep data without the constraints of clinical scoring.
• To validate the framework's capability by predicting subject age and estimating biological age.

Main Methods:

• A hierarchical attention network architecture was developed for deep learning-based sleep analysis.
• The network was pre-trained on 5-minute data epochs and fine-tuned for whole-night polysomnography recordings.
• The model was trained on 511 recordings and validated on 146 subjects (ages 6-88) from the Cleveland Family study.

Main Results:

• The deep learning framework achieved a mean absolute error of 7.36 years in age prediction.
• The model demonstrated a strong correlation of 0.857 between predicted and true chronological age.
• The framework successfully processed whole-night polysomnograms, indicating its potential for generalization.

Conclusions:

• An end-to-end deep learning framework can effectively analyze polysomnograms for age prediction.
• This framework offers a potential tool for estimating biological age from sleep data.
• The approach is expected to generalize to learning other subject-specific labels, including sleep disorders.