Benchmark on a large cohort for sleep-wake classification with machine learning techniques.

Area of Science:

• Sleep Science
• Biomedical Engineering
• Data Science

Background:

• Polysomnography (PSG) is the gold standard for sleep measurement but is costly.
• Actigraphy offers a cheaper alternative but has been limited by smaller datasets.
• Previous large-scale comparisons of sleep-wake scoring algorithms are lacking.

Purpose of the Study:

• To process and utilize the large-scale Multi-Ethnic Study of Atherosclerosis (MESA) Sleep dataset, synchronizing polysomnography (PSG) and actigraphy data.
• To systematically compare existing sleep-wake stage detection algorithms, including traditional and machine learning approaches.
• To foster the development of novel sleep-wake scoring algorithms by providing a robust, large-scale dataset.

Main Methods:

• Processed synchronized PSG and actigraphy data from the MESA Sleep study, a dataset significantly larger than previous ones.
• Implemented and compared traditional algorithms and state-of-the-art machine learning methods for sleep-wake stage scoring.
• Evaluated algorithm performance using metrics such as accuracy and F1 score.

Main Results:

• Identified traditional algorithms outperforming the MESA Sleep actigraphy device's native algorithm.
• Machine learning algorithms demonstrated superior performance compared to the device's native algorithm, achieving accuracy comparable to human annotation.
• Specific deep-learning architectures (CNN, LSTM) achieved significantly higher accuracy scores than other methods.

Conclusions:

• The MESA Sleep dataset is a valuable resource for advancing sleep-wake scoring algorithm research.
• Machine learning approaches, particularly deep learning architectures like CNN and LSTM, show significant promise for accurate sleep-wake stage detection.
• Future research can leverage this large dataset to develop and validate improved sleep scoring algorithms.