DREAMT: Dataset for Real-time sleep stage EstimAtion using Multisensor wearable Technology

Abstract

Sleep is an intrinsic part of human life, and recent advancements in wearable technology and machine learning have promised continuous and non-invasive methods of tracking sleep health and patterns, providing an important facet to a more holistic understanding of well-being. However, it is still challenging to achieve consistent and reliable real-time estimates of sleep stages using only smartwatches. This is especially true for individuals with irregular sleep patterns or sleep disorders. A major contributing factor is the distinct lack of publicly accessible, large-scale datasets that allow researchers and engineers to validate their wearable sleep staging algorithms against a population with diverse sleep patterns. Here, we present DREAMT, Dataset for Real-time sleep stage EstimAtion using Multisensor wearable Technology, a new dataset collected from 100 participants, which includes high-resolution signals from a smartwatch, expert sleep technician-annotated sleep stage labels, and clinical metadata related to sleep health and disorders.

Background

Human health is greatly impacted by the overall quantity and quality of our sleep ([1, 2, 3, 4]). Shorter sleep duration and poor sleep quality are associated with increased prevalence and severity of chronic diseases and mental health conditions independent of socioeconomic and behavioral factors or sleep disorders ([5, 6]). About 70 million U.S. adults are reported to have some type of sleep disorder ([7]), such as Obstructive Sleep Apnea (OSA), which impedes proper ventilation during sleep, leading to non-restorative sleep and other detrimental symptoms. There is a clear and urgent need for continuous sleep monitoring to help patients with atypical sleep in real-world settings and to enable timely and personalized interventions ([8]). The current gold-standard method of sleep assessment is polysomnography (PSG), which involves multiple simultaneous gold-standard sensing modalities ([8, 9]), but is also labor-intensive, costly, obtrusive, and takes place in non-natural sleep settings. Small, low-cost, wearable, and highly portable sensors can provide the opportunity for a new mode of sleep assessment ([10]). Unfortunately, many of the existing sleep stage estimation algorithms that use signals from wearable sensors lack reliability and accuracy. The challenge of developing such algorithms that work for everyone is worsened by the limited availability of open-source or public datasets for sleep stage estimation using signals from wearable devices, as well as a lack of publicly available sleep stage estimation algorithms to benchmark novel methods against.

Methods

A total of 100 unique participants were recruited from the Duke University Health System (DUHS) Sleep Disorder Lab to participate in the study between May 2022 and September 2022. Upon arrival at the sleep lab, each participant received comprehensive information about the study according to IRB (#Pro00108961). Written informed consent was obtained for all study participants.

The sleep study protocol is designed to detect and monitor patients’ apnea events during sleep. Patients are asked to abstain from caffeine after noon on the day of PSG recording to ensure standard protocol. There is no explicit monitoring of caffeine intake on the arrival of the participant in the sleep lab.

Each participant checked in at approximately 8 PM on the scheduled day of their appointment. They stayed in a hotel-like room for a night at the sleep lab, in different rooms on the same floor. After they checked in for their study at the sleep lab, the clinician would ask for their willingness to join our data collection. Due to the limited number of devices we possessed, we were only able to recruit at most 4 participants each night. There were no selection criteria for the potential participants since we were aiming for people with sleep apnea. The Empatica E4 wristband (Empatica Inc., Milano, Italy; Release Date: NA; Software Version: Summer 2019) was placed on the participant's left wrist immediately after the consent form was signed at the participant’s arrival. Overnight PSG was performed using the Nihon Khoden Polysmith 1004 (version 11) Data Management System (DMS). The PSG data collection was started at approximately 11 PM by the nurse. The recording continued throughout the night to monitor the participant’s sleep condition and ended when the participant awakened around 6 AM the next morning naturally. The E4 device was also deactivated at the time of awakening.

The E4 device collected 6 raw signals, Blood Volume Pulse (BVP) derived from the photoplethysmography (PPG) sensor, Accelerometry (ACC) in 3 axes, Electrodermal Activity (EDA), and Skin Temperature (TEMP). Heart Rate (HR) and Inter-beat Interval (IBI) were estimated from the raw BVP. We also retrieved the technician-annotated sleep labels derived from the PSG data. All the data were resampled to the highest sampling frequency (64 Hz) and were joined together by the timestamp. The actual timestamp was time-shifted and started with 0 to preserve privacy. There was no data preprocessing performed on the raw data other than resampling.

The average age of the participants was 56.2 ± 16.6 years, BMI: 33.7 ± 8.6 kg/m², OAHI: 19.4 ± 27.5 /h, AHI: 22.1 ± 28.7 /h (AHI < 5/h is healthy). Among all participants, 68 were obese or severely obese (BMI ≥ 30 kg/m²). Among the 23 participants who had severe OSA (AHI > 30 /h), 17 were obese or severely obese. More detailed information about the participant was included in participant_info.csv.

Data Description

The time-aligned dataset consists of six E4 raw signals (BVP, ACC_X, ACC_Y, ACC_Z, EDA, TEMP), two derived signals (HR and IBI), the sleep-stage label (REM Sleep, Non-REM Sleep, Wake), and the true timestamp of every epoch.

• TIMESTAMP [s] (64 Hz): The timestamp shifted and started with 0, with a frequency of 64 Hz.
• BVP (64 Hz): Blood volume pulse derived from the photoplethysmography (PPG) sensor.
• IBI [ms]: Inter-beat interval is the time interval between individual beats of the heart, derived from the photoplethysmography (PPG) sensor.
• EDA [μS] (4 Hz): Electrodermal activity from the galvanic skin response sensor.
• TEMP [°C] (4 Hz): Skin temperature from the infrared thermopile sensor.
• ACC (32 Hz): Triaxial accelerometry with each axis named ACC_X, ACC_Y, and ACC_Z.
• HR [bpm] (1 Hz): Heart rate is estimated from the BVP signal.
• Sleep_Stage: The technician-annotated sleep labels derived from PSG are recorded every 30 seconds.
  • P: Preparation stage, stages before the PSG recording starts
  • W: Wake stage
  • N1: Non-rapid eye-movement (NREM) stage 1
  • N2: Non-rapid eye-movement (NREM) stage 2
  • N3: Non-rapid eye-movement (NREM) stage 3
  • R: Rapid eye movement (REM)
  • Missing: No sleep stage labeled on PSG
    • The occurrence of the "Missing" label is extremely low. Significant missingness has been found in only 2 participants, who had their PSG re-setup during the overnight study, which resulted in 15 minutes of consecutive missing labels each. We also found one epoch with "Missing" label each in 4 other participants.

A more detailed explanation of the data and acronym used in our dataset were included in metadata.txt.

Usage Notes

In addition to the continued development and testing of wearable sleep stage estimation models, this dataset will also enable the extraction of new knowledge through unsupervised learning methods. For example, researchers can examine how severity of OSA might affect sleep patterns across different age ranges and gender, and uncover new subgroups of patients who would benefit from different treatment paths. Also, predictive models can be developed to predict key health outcomes, such as AHI, from wearables data alone, providing new opportunities for mobile population health applications and interventions.

Ethics

Study information was provided by the DUHS Sleep Disorders Lab care team following the ethics protocols established by the Duke Health IRB, including written informed consent (IRB #Pro00108961) with explicit permission to share de-identified data. In the publicly-available data, all direct identifiers are removed and all timestamps are time shifted to protect participant identities.

Acknowledgements

We acknowledge the support of Duke Sleep Disorders Lab with recruiting participants and providing the sleep stage annotations.

Conflicts of Interest

The author(s) have no conflicts of interest to declare.

References

1. Anil Natesan Rama, S. Charles Cho, and Clete A. Kushida. Normal Human Sleep. In Sleep: A Comprehensive Handbook, pages 1–9. John Wiley & Sons, Ltd, 2005. ISBN 978-0-471-75172-4. doi: 10.1002/0471751723.ch1. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/0471751723.ch1. Section: 1 _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/0471751723.ch1.
2. Jane E. Ferrie, Martin J. Shipley, Tasnime N. Akbaraly, Michael G. Marmot, Mika Kivimäki, and Archana Singh-Manoux. Change in sleep duration and cognitive function: findings from the Whitehall II Study. Sleep, 34(5):565–573, May 2011. ISSN 1550-9109. doi: 10.1093/sleep/34.5.565.
3. Jane E. Ferrie, Mika Kivimäki, Tasnime N. Akbaraly, Archana Singh-Manoux, Michelle A.Miller, David Gimeno, Meena Kumari, George Davey Smith, and Martin J. Shipley. Associations Between Change in Sleep Duration and Inflammation: Findings on C-reactive Protein and Interleukin 6 in the Whitehall II Study. American Journal of Epidemiology, 178(6):956–961, September 2013. ISSN 0002-9262. doi: 10.1093/aje/kwt072. URL https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3817449/.
4. Leah A. Irish, Christopher E. Kline, Heather E. Gunn, Daniel J. Buysse, and Martica H. Hall.The Role of Sleep Hygiene in Promoting Public Health: A Review of Empirical Evidence. Sleep medicine reviews, 22:23–36, August 2015. ISSN 1087-0792. doi: 10.1016/j.smrv.2014.10.001. URL https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4400203/.
5. Sari Stenholm, Jenny Head, Mika Kivimäki, Linda L. Magnusson Hanson, Jaana Pentti, Naja H. Rod, Alice J. Clark, Tuula Oksanen, Hugo Westerlund, and Jussi Vahtera. Sleep Duration and Sleep Disturbances as Predictors of Healthy and Chronic Disease-Free Life Expectancy Between Ages 50 and 75: A Pooled Analysis of Three Cohorts. The Journals of Gerontology.Series A, Biological Sciences and Medical Sciences, 74(2):204–210, January 2019. ISSN386 1758-535X. doi: 10.1093/gerona/gly016.
6. Maria Ruiz-Castell, Tatjana T. Makovski, Valéry Bocquet, and Saverio Stranges. Sleep duration and multimorbidity in Luxembourg: results from the European Health Examination Survey in Luxembourg, 2013-2015. BMJ open, 9(8):e026942, August 2019. ISSN 2044-6055. doi:10.1136/bmjopen-2018-026942.
7. Common Sleep Disorders: Symptoms, Causes & Treatment, . URL https://my.clevelandclinic.org/health/articles/11429-common-sleep-disorders.
8. Ignacio Perez-Pozuelo, Bing Zhai, Joao Palotti, Raghvendra Mall, Michaël Aupetit, Juan M. Garcia-Gomez, Shahrad Taheri, Yu Guan, and Luis Fernandez-Luque. The future of sleep health: a data-driven revolution in sleep science and medicine. npj Digital Medicine, 3(1):1–15, March 2020. ISSN 2398-6352. doi: 10.1038/s41746-020-0244-4. URL https://www.nature.com/articles/s41746-020-0244-4. Number: 1 Publisher: Nature Publishing Group.
9. Behrouz Jafari and Vahid Mohsenin. Polysomnography. Clinics in Chest Medicine, 31(2):287–297, June 2010. ISSN 1557-8216. doi: 10.1016/j.ccm.2010.02.005.
10. Fred Shaffer and J. P. Ginsberg. An Overview of Heart Rate Variability Metrics and Norms. Frontiers in Public Health, 5:258, September 2017. ISSN 2296-2565. doi: 10.3389/fpubh.2017.00258. URL https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5624990/.