Automatic Actigraphy and Polysomnography Sleep Scoring using Deep Learning

Abstract

The utilization of deep learning models for sleep scoring has become an increasingly promising area of research due to their potential to automate and enhance the accuracy of this crucial task. Sleep scoring involves categorizing a patient's polysomnography (PSG) data into different sleep stages, which plays a vital role in diagnosing sleep disorders and understanding an individual's sleep patterns. In this study, two significant sources of data were employed: actigraphy and PSG recordings. Actigraphy, a non-invasive method, captures physical activity and light exposure, enabling sleep/wake prediction. PSG, on the other hand, incorporates various physiological signals, such as EEG, ECG, and EOG recordings, providing comprehensive insights into brain activity, cardiac activity, and eye movements during sleep. To address the complexity of sleep scoring and improve accuracy, three deep learning architectures were chosen for evaluation: Convolutional – Long Short-Term Memory (CNNLSTM), Extreme Gradient Boosting (XGBoost), and LSTM. These models were assessed on a dataset comprising 109 subjects for actigraphy sleep/wake prediction and 30 subjects for PSG sleep staging. Each subject's dataset consisted of five nights of sleep data, offering diverse samples. The integration of actigraphy and PSG data proved to be a valuable strategy, providing a more comprehensive understanding of an individual's sleep architecture. By utilizing the power of deep learning models and incorporating multi-modal data, clinicians and researchers can significantly improve sleep disorder diagnosis and treatment. The potential for automating the sleep scoring process promises to enhance the efficiency of sleep studies, allowing healthcare professionals to focus on tailored treatment plans and better patient care. As the availability of large-scale sleep datasets and computational resources continues to grow, the future of sleep scoring with deep learning models holds great promise. With ongoing research and advancements, these models have the potential to become indispensable tools in sleep medicine, empowering healthcare providers to optimize sleep health and overall well-being for their patients.