Automatic sleep stage classification with deep residual networks in a mixed-cohort setting

Area of Science:

• Sleep Medicine
• Artificial Intelligence
• Computational Neuroscience

Background:

• Manual sleep stage scoring suffers from subjectivity and low reliability.
• Existing automatic systems often lack generalizability due to reliance on limited datasets.

Purpose of the Study:

• To investigate the generalizability of a novel deep neural network for sleep stage scoring across large-scale cohorts.
• To assess the impact of data diversity and volume on model performance.

Main Methods:

• Developed a deep neural network using 15,684 polysomnography studies from five distinct cohorts.
• Evaluated model performance under various training scenarios, including varying timescales, cohort sizes, and data source combinations.

Main Results:

• Classification accuracy increased with larger fractions of training data, reaching 0.869 ± 0.064 with 100% data.
• Performance improved with more data sources, achieving 0.821 ± 0.085 with four cohorts.
• Generalization varied across different cohorts.

Conclusions:

• Deep learning models for automatic sleep stage scoring require extensive data from multiple sources for robust generalization.
• Availability of public datasets is crucial for benchmarking and advancing future research in automated sleep analysis.