Personalized Sleep Parameters Estimation from Actigraphy: A Machine Learning Approach.

Area of Science:

• Biomedical Engineering
• Sleep Science
• Machine Learning

Background:

• Polysomnography (PSG) is the gold standard for sleep measurement but is costly and obtrusive.
• Actigraphy offers a low-cost, unobtrusive alternative for sleep-wake state prediction.
• Generalized models using population data have limitations in capturing individual sleep variability.

Purpose of the Study:

• To validate personalized machine learning models for sleep-wake state prediction using individual actigraphy data.
• To assess if personalized models improve the accuracy of nightly sleep parameter estimation compared to generalized models.

Main Methods:

• Trained and tested five personalized machine learning models and their generalized counterparts on data from 54 participants.
• Compared model performance against concurrent polysomnography (PSG) using machine learning experiments and statistical analyses.

Main Results:

• Personalized models significantly outperformed generalized models in estimating key sleep parameters (total sleep time, wake after sleep onset, sleep efficiency, number of awakenings).
• Estimates from personalized models showed statistically non-significant differences compared to PSG-derived values.
• Specific personalized models (regularized logistic regression, random forest, adaptive boosting, extreme gradient boosting) demonstrated superior performance.

Conclusions:

• Personalized machine learning models analyzing actigraphy data are superior to generalized models for sleep parameter estimation.
• These personalized models achieve results indistinguishable from PSG, supporting their use in sleep health studies.
• Personalized actigraphy models show potential for screening sleep disorders.