A Multi-Night Instantaneous Heart Rate and Accelerometry Dataset with EEG Sleep Stage Labels

@article{PhysioNet-bidsleep-dataset-1.0.0,
  author = {Song, Tzu-An},
  title = {{A Multi-Night Instantaneous Heart Rate and Accelerometry Dataset with EEG Sleep Stage Labels}},
  journal = {{PhysioNet}},
  year = {2026},
  month = may,
  note = {Version 1.0.0},
  doi = {10.13026/a0sy-7t69},
  url = {https://doi.org/10.13026/a0sy-7t69}
}

Abstract

We collected sleep data from 47 healthy adult volunteers with no history of sleep disorders, recruited from the local community through study advertisement flyers. The study protocol was approved by the University of Massachusetts Lowell Institutional Review Board. Participants were selected without regard for gender or ethnicity and provided written informed consent. Each participant simultaneously wore a Dreem 2 Headband and an Apple Watch for up to seven consecutive nights, yielding a total of 253 nights of data.

The dataset contains multi-night recordings of instantaneous heart rate and 3-axis accelerometry from the Apple Watch, along with sleep stage labels derived from Dreem 2 EEG and scored according to the AASM standard (Wake, N1, N2, N3, REM). This resource supports the development and validation of wearable-based sleep staging methods, multimodal modeling of heart rate and motion signals, and investigations that benefit from multi-night sleep data.

Background

Sleep is a fundamental biological process that is closely linked to physical and mental health [1, 2]. Polysomnography (PSG) is the traditional way to accurately describe sleep stages. It records many physiological signals, such as electroencephalography (EEG), electrocardiography (ECG), electrooculography (EOG), and electromyography (EMG) [3]. While PSG is considered the gold standard, it is costly, resource-intensive, and difficult to scale for large, multi-night studies [4].

Wearable devices such as smartwatches offer an accessible alternative for large-scale and longitudinal monitoring of sleep [5]. They commonly provide photoplethysmography (PPG)-derived heart rate measures and wrist-based accelerometry, which can be leveraged to infer sleep stages. However, consumer devices often report sleep summaries through proprietary “black box” algorithms, limiting transparency and reproducibility in research.

The dataset described here addresses this gap by providing multi-night, open-access recordings from the Apple Watch aligned with EEG-based reference labels from the Dreem 2 headband. The dataset pairs consumer smartwatch data, including instantaneous heart rate and 3-axis accelerometry, with sleep stage annotations scored according to the American Academy of Sleep Medicine (AASM) standard [3].

By making these data openly available, this resource enables the development and validation of sleep staging algorithms using wearable signals, supports investigations into the reliability of consumer wearables for sleep research, and facilitates studies on generalizability across devices and across multiple nights of recordings. Prior work has demonstrated the feasibility of sleep staging using heart rate and accelerometry, including our previous studies [6, 7].

Methods

Participants

The dataset includes multi-night sleep recordings from 47 healthy adult participants. Each participant contributed up to seven nights of data, resulting in a total of 253 nights (with the exact number varying by individual).

Participants were screened prior to enrollment using self-reported health questionnaires and study-specific inclusion criteria. Individuals were included if they were generally healthy adults without known sleep or cardiovascular conditions that could substantially affect sleep physiology.

Participants were excluded if they reported:

• Severe untreated sleep disorders, defined as a prior clinical diagnosis of conditions such as obstructive sleep apnea (OSA), insomnia disorder, narcolepsy, or other sleep disorders that were not currently treated or managed.
• Major cardiovascular conditions, defined as a history of significant heart-related diseases such as arrhythmia, heart failure, coronary artery disease, or other clinically diagnosed cardiovascular conditions requiring ongoing medical management.

Screening was based on self-report and study intake procedures; no additional clinical diagnostic testing was performed as part of this study.

All procedures were approved by the institutional review board, and written informed consent was obtained from all participants.

Devices and Data Collection

• Apple Watch (consumer-grade device): A custom-developed WatchOS application, BIDSleep, was used to continuously record instantaneous heart rate (IHR) and three-axis accelerometry.
• Dreem-2 EEG headband (research-grade device): Served as the reference for sleep staging. Sleep stage labels (Wake, N1, N2, N3, REM) were generated according to the American Academy of Sleep Medicine (AASM) standard, with 30-second epoch resolution.

Procedure

Participants wore both the Apple Watch and Dreem 2 EEG headband at home during sleep. To ensure accurate alignment of sleep staging data with other physiological signals, data from both devices were recorded using the same Apple iPhone. This approach provided a common time reference, minimizing potential misalignment issues between the two devices.

Reference Labels

Dreem 2 provided sleep stage annotations in 30-second epochs. All Dreem EEG signals and their corresponding auto-scored hypnograms were reviewed by a single trained sleep expert to verify signal quality and resolve artifacts prior to analysis. As only one scorer was involved, no inter-scorer adjudication was performed.

The review included signal quality inspection and manual epoch-by-epoch verification following AASM guidelines, with attention to ambiguous transitions and physiologically implausible stage sequences using contextual information from adjacent epochs. Both the original automated labels and the expert-corrected labels are included in the dataset.

Data Description

Dataset Organization and Contents

The dataset is organized into 47 subject-level folders, with each folder containing between 3 and 7 nights of raw data. For each night, three files are provided: motion.csv, hr.csv, and labels.mat.

• motion.csv: Three-axis (x, y, z) accelerometer data recorded by the Apple Watch. Each row includes a timestamp in Unix time (seconds, with sub-second precision) indicating when the measurement was recorded.
• hr.csv: Instantaneous heart rate (IHR) values derived from the Apple Watch’s PPG sensor via HealthKit. IHR is measured in beats per minute (bpm) at specific time points, with an approximate sampling rate of 0.2 Hz. Each row includes a timestamp in Unix time (seconds, with sub-second precision).
• labels.mat: Contains three key variables:
  • recStart: Timestamp marking the start of the recording.
  • dreem_label: Automated sleep stage annotations from the Dreem device, encoded as integers (0 = Wake, 1 = N1, 2 = N2, 3 = N3, 4 = REM, 5 = Unknown).
  • expert_label: Manual sleep stage annotations by a sleep expert using the same coding scheme.

Timestamp Alignment

The hr.csv and motion.csv files contain timestamps represented as Unix time in seconds (floating-point values) with sub-second precision. These timestamps indicate absolute time referenced to the Unix epoch (January 1, 1970, UTC).

In labels.mat, the variable recStart denotes the start time of the recording. The original timestamp is obtained in Unix time and converted to a human-readable format in U.S. Eastern Time (ET) for storage.

To align physiological signals with sleep stage labels:

1. Convert recStart to Unix time (if needed).
2. Each sleep stage label corresponds to a 30-second epoch relative to recStart:
   • 1st epoch:
     [recStart, recStart + 30 s)
   • k-th epoch:
     [recStart + 30 × (k − 1), recStart + 30 × k)
3. For any timestamp t in hr.csv or motion.csv, the corresponding sleep stage index k can be computed as:

   k = floor((t − recStart) / 30) + 1

This alignment enables direct mapping of heart rate and accelerometry samples to their corresponding sleep stage epochs.

Example (Python):

import datetime
import pytz

dt = datetime.datetime.strptime("2021-12-02 23:11:25", "%Y-%m-%d %H:%M:%S")
dt = pytz.timezone("US/Eastern").localize(dt)
unix_time = dt.timestamp()

Usage Notes

This dataset can be used to develop and evaluate methods for wearable-based sleep staging, as well as to study how heart rate and movement change across multiple nights of sleep. Data are provided in two forms: raw instantaneous heart rate and 3-axis accelerometry signals from the Apple Watch, and 30-second sleep stage labels from the Dreem 2 EEG headband.

We recommend performing analyses at the subject level (e.g., cross-validation by participant) rather than mixing data across nights to avoid data leakage. While the Dreem 2 headband provides reliable EEG-based staging, it is not equivalent to full polysomnography, and minor discrepancies may occur in transitional epochs. In our work, models were trained on a per-night basis; however, we encourage future research to explore approaches that leverage multi-night information within individuals, as this dataset is well-suited to support such analyses.

All subject-night folders in this dataset are complete and contain the expected three files:
motion.csv, hr.csv, and labels.mat.

These modalities are available for all subject-night folders.

Release Notes

Version 1.0.0: Initial release of the dataset.

Ethics

The study protocol was reviewed and approved by the University of Massachusetts Lowell Institutional Review Board. All participants provided written informed consent prior to enrollment. All data were de-identified in accordance with institutional and ethical guidelines before public release.

Acknowledgements

This work was supported in part by the National Institutes of Health grants R21AG068890, R01AG082354, and P30AG073107, and the Dreem Jury’s Prize. We also extend our gratitude to all the anonymous participants for their time and effort in recording their sleep data, which made this research possible.

Conflicts of Interest

The authors declare no conflicts of interest.

References

1. Malhotra A, Loscalzo J. Sleep and cardiovascular disease: an overview. Prog Cardiovasc Dis. 2009 Jan;51:279–284.
2. Shi L, Chen SJ, Ma MY, Bao YP, Han Y, Wang YM, Shi J, Vitiello MV, Lu L. Sleep disturbances increase the risk of dementia: A systematic review and meta-analysis. Sleep Med Rev. 2018 Aug;40:4–16.
3. Moser D, Anderer P, Gruber G, Parapatics S, Loretz E, Boeck M, Kloesch G, Heller E, Schmidt A, Danker-Hopfe H, Saletu B, Zeitlhofer J, Dorffner G. Sleep classification according to AASM and Rechtschaffen & Kales: effects on sleep scoring parameters. Sleep. 2009 Feb;32:139–149.
4. Kim RD, Kapur VK, Redline-Bruch J, Rueschman M, Auckley DH, Benca RM, Foldvary-Schafer NR, Iber C, Zee PC, Rosen CL, Redline S, Ramsey SD. An economic evaluation of home versus laboratory-based diagnosis of obstructive sleep apnea. Sleep. 2015 Jul;38:1027–1037.
5. Shelgikar AV, Anderson PF, Stephens MR. Sleep tracking, wearable technology, and opportunities for research and clinical care. Chest. 2016 Sep;150:732–743.
6. Song TA, Chowdhury SR, Malekzadeh M, Harrison S, Hoge TB, Redline S, Stone KL, Saxena R, Purcell SM, Dutta J. AI-driven sleep staging from actigraphy and heart rate. PLoS One. 2023 May;18:1–29.
7. Song TA, Zhang Y, Zhou Z, Hou L, Malekzadeh M, Behzad A, Dutta J. AI-driven sleep staging using instantaneous heart rate and accelerometry: insights from an Apple Watch study. IEEE Trans Biomed Eng. 2026 Apr;73(4):1596–1608. doi:10.1109/TBME.2025.3612158.