Personalized automatic sleep staging with single-night data: a pilot study with Kullback-Leibler divergence...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Individual variations in brain waves necessitate personalized approaches for accurate automatic sleep staging.
• Longitudinal sleep monitoring relies on robust algorithms capable of adapting to individual characteristics.
• Training models with limited single-night sleep data presents a significant challenge.

Purpose of the Study:

• To enhance automatic sleep staging for longitudinal monitoring through algorithm personalization.
• To develop a method for adapting sleep staging algorithms using data from the first night of sleep.
• To address the challenge of limited data for model training via transfer learning.

Main Methods:

• Proposed a Kullback-Leibler (KL) divergence regularized transfer learning approach.
• Utilized a pretrained subject-independent model (SeqSleepNet) as a baseline.
• Fine-tuned the model with single-night personalization data, incorporating KL divergence into the loss function to prevent overfitting.

Main Results:

• Demonstrated successful sleep staging personalization using single-night data with KL-divergence regularization.
• Achieved an average personalized sleep staging accuracy of 79.6%, Cohen's kappa of 0.706, and macro F1-score of 73.0%.
• Reported sensitivity of 71.8% and specificity of 94.2% on the Sleep-EDF Expanded database.

Conclusions:

• The proposed KL-divergence regularized transfer learning approach is effective for single-night sleep staging personalization.
• The method significantly improves accuracy compared to baseline and non-regularized personalization methods.
• The approach is robust against overfitting, enabling reliable longitudinal sleep monitoring.