Enhancing ECG-Based Sleep Staging and Overcoming Challenges in Wearable Sleep Monitoring

  • 10



Name of the Speaker: Mr. JOSHI VAIBHAV SHRIPAD (EE19S039)
Guide: Dr. Mohanasankar S
Venue/Online meeting link:
Date/Time: April 10th 2023 (Monday), at 4.00 PM

Sleep is considered a fundamental physiological process that is deemed essential for maintaining good health and overall well-being. In recent years, wearable devices have gained popularity due to their non-intrusive and convenient approach to monitoring various aspects of human physiology, including sleep. Accurate and reliable sleep monitoring is deemed crucial for understanding and treating sleep disorders that can significantly improve the quality of life for millions of people.

However, traditional polysomnography, which is the gold standard for sleep monitoring, is highly inconvenient and requires multiple signals. Although electroencephalogram (EEG) is the best signal for single-channel sleep staging, it is expensive and requires a clinical setup that can be intrusive in nature, making it unsuitable for point-of-care settings. On the other hand, electrocardiogram (ECG) is less intrusive and more suitable for use in point-of-care settings, but its performance in sleep staging is subpar compared to EEG.

To address this issue, a deep learning-based cross-modality Knowledge Distillation (KD) framework to transfer knowledge from EEG to ECG has been proposed here. The proposed model was tested on the Montreal Archive of Sleep Studies (MASS) dataset of 200 subjects, resulting in a 13.40 % and 14.30 % improvement in weighted F1-score for 3-class and 4-class sleep staging, respectively. These results demonstrate the feasibility of using KD for enhancing ECG-based sleep staging.

While wearable devices have made sleep monitoring more convenient and non-intrusive, the accuracy of these devices is still a concern. In this study, the challenges of wearable-based sleep staging were addressed using machine learning and statistical exploration approaches on the wearable device product "Repose". However, the validation of ground-truth sleep stages and sleep onset identification relies on the annotators' observation. A sleep onset detection algorithm was developed using sleep position, movement data and statistical HRV features, including mean, variation, and standard deviation of heart rate during sleep. The performance improved by 44.1 % over the previous sleep onset detection algorithm. Additionally, machine learning algorithm XG Boost was tested on the MASS data with the aim of validating the algorithm and using it on the Repose data. However, it was found that the statistical approach outperforms the machine learning approach.

These results open up new avenues for research in sleep monitoring and pave the way for developing more accessible and reliable solutions for sleep monitoring.