Refining sleep staging accuracy: transfer learning coupled with scorability models.

Area of Science:

• Artificial Intelligence in Medicine
• Sleep Medicine
• Biomedical Signal Processing

Background:

• Automated sleep staging using polysomnography (PSG) is crucial for diagnosing sleep disorders.
• Existing AI models often struggle with reduced montage or limited data, impacting accuracy.
• Human expert agreement in sleep staging provides a benchmark for automated systems.

Purpose of the Study:

• To enhance sleep staging accuracy using transfer learning (TL) to meet or exceed human inter-expert agreement.
• To develop a scorability model for evaluating the quality and trustworthiness of automated sleep staging results.
• To validate AI performance against human experts using a Korean dataset (KoGES).

Main Methods:

• A deep neural network was initially trained on a large US PSG dataset.
• Transfer learning (TL) was applied to calibrate the model using a reduced montage and limited samples from the Korean Genome and Epidemiology Study (KoGES).
• Model performance was compared to inter-expert reliability among three human scorers, and a scorability assessment was developed.

Main Results:

• The base model showed moderate agreement (κ = 0.55), lower than inter-expert agreement (κ = 0.62).
• TL calibration with targeted sampling improved performance, with models exceeding inter-expert agreement (κ = 0.70).
• The scorability assessment moderately predicted model-expert agreement (R² = 0.42), indicating higher agreement for recordings with higher scores.

Conclusions:

• Targeted TL significantly enhances automated sleep staging performance on atypical montages, surpassing human expert agreement.
• The scorability assessment offers a reliable method for quality control in automated sleep analysis.
• This AI-driven approach advances automated sleep analysis, showing potential to exceed human performance in clinical settings.