Benchmarking of Sleep/Wake Detection Algorithms on a Large Cohort using Actigraphy, HRV, and Respiration Information

Krauß D, Richer R, Küderle A, Beilner J, Rohleder N, Eskofier B (2022)

Publication Language: English

Publication Type: Conference contribution, Abstract of a poster

Publication year: 2022

Series: IEEE-EMBS International Conference on Biomedical and Health Informatics

Event location: Ioaninna

Abstract

I. INTRODUCTION

Insufficient sleep quality is directly linked to a series of physical and physiological diseases [1]. Therefore, reliable sleep monitoring is essential for the prevention, diagnosis, and treatment of such. As sleep laboratories are very cost- and resource-prohibitive, wearable sensors are a promising alternative for unobtrusive sleep monitoring at home. During sleep, body movements decrease compared to a wakeful state [2]. In addition, cardiac and respiratory activity changes during sleep [3]. Current systems are mostly based on wrist movement, typically assessed using actigraphy (ACT), for unobtrusive sleep/wake detection [4]. However, movement-based systems tend to overestimate sleep due to a lack of movement shortly before falling asleep or in short periods of wakefulness. Previous research showed promising improvements in sleep/wake detection by combining ACT with cardiac and respiratory information such as heart rate variability (HRV) and respiration rate variability (RRV) [5]. However, this was only evaluated on small cohorts and not in large-scale studies. For that reason, this work aims to systematically compare ACT-based sleep detection with multimodal approaches combining ACT and HRV by benchmarking different state-of-the-art machine- and deep learning algorithms on a large-scale dataset. In particular, we investigate whether the classification performance can be further improved by including respiratory information into machine learning models.

II. METHODS

The data used in this work were collected in a sleep study of 2,237 participants, which contains ACT and polysomnography (PSG). PSG was used as ground truth for sleep/wake phases as well as to extract HRV and RRV from electrocardiography and respiratory induction plethysmography respectively [7]. In total, 370 ACT features and 30 HRV features were extracted according to Zhai et al. [6]. In addition, 62 RRV features were extracted using the Neurokit2 library [8]. To find the best set of hyperparameters, a grid search with embedded 5-fold cross-validation was performed over a defined search space.

III. RESULTS & DISCUSSION

Our results show that including RRV features in the classification algorithms significantly improved the key metrics of assessing sleep/wake detection performance (Fig 1). The best-performing algorithm to discriminate between sleep and wake phases was a Multi-Layer Perceptron with an accuracy of 85.1±8.5%. In particular, specificity, which is a good marker for assessing the overprediction of sleep, showed a strong increase in performance after adding RRV (63.5±22.3% vs. 72.4±17.2%). Our findings underscore the potential of including respiratory information, which can also be extracted from wearable sensors, to improve sleep/wake detection algorithms and, thus, help to transfer sleep laboratories into a home monitoring environment.

Authors with CRIS profile

Additional Organisation(s)

Related research project(s)

How to cite

APA:

Krauß, D., Richer, R., Küderle, A., Beilner, J., Rohleder, N., & Eskofier, B. (2022, October). Benchmarking of Sleep/Wake Detection Algorithms on a Large Cohort using Actigraphy, HRV, and Respiration Information. Poster presentation at IEEE-EMBS International Conference on Biomedical and Health Informatics, Ioaninna, GR.

MLA:

Krauß, Daniel, et al. "Benchmarking of Sleep/Wake Detection Algorithms on a Large Cohort using Actigraphy, HRV, and Respiration Information." Presented at IEEE-EMBS International Conference on Biomedical and Health Informatics, Ioaninna 2022.

BibTeX: Download