Auricle: A Multimodal Wearable Sensing Dataset for Cardiovascular and Hemodynamic Assessment

@article{PhysioNet-auricle-1.0.0,
  author = {Christofferson, Kenneth and Bhalla, Sejal and Kleinknecht, Thomas and Varickanickal, Joann and Kim, Ben and {De Luca}, Enza and Simard, Anne and Ross, Heather and Cafazzo, Joseph and Mariakakis, Alex},
  title = {{Auricle: A Multimodal Wearable Sensing Dataset for Cardiovascular and Hemodynamic Assessment}},
  journal = {{PhysioNet}},
  year = {2026},
  month = oct,
  note = {Version 1.0.0},
  doi = {10.13026/swme-vp69},
  url = {https://doi.org/10.13026/swme-vp69}
}

@article{pollard_physionet_2026, title = {{PhysioNet} as a global platform for biomedical research}, volume = {1}, issn = {3005-0693}, url = {https://doi.org/10.1038/s44360-026-00096-z}, doi = {10.1038/s44360-026-00096-z}, abstract = {Twenty-five years after its first public repository was established, the PhysioNet resource for biomedical and clinical data has enabled tens of thousands of research papers. Further efforts are ongoing to advance secure data sharing, community engagement and scalable research infrastructure.}, number = {8}, journal = {Nature Health}, author = {Pollard, Tom and Moody, Benjamin E. and Lehman, Li-wei H. and Gow, Brian J. and Fernandes, Chrystinne and Xie, Chen and Johnson, Alistair and Mark, Roger G. and Heldt, Thomas}, month = aug, year = {2026}, pages = {792--795},}

Abstract

Wearable cardiovascular monitoring aims to transition cardiovascular care from episodic clinical visits to continuous, free-living assessment. Many proposed systems use smartwatches and smart rings to capture varying combinations of photoplethysmography (PPG), electrocardiography (ECG), and seismocardiography (SCG) signals. More recently, earable devices like wireless earbuds have been leveraged for similar sensing applications; beyond being equipped with PPG, they can also detect in-ear cardiac audio using the embedded microphones required for active noise cancellation. Regardless of form factor or signal modality, wearable cardiovascular monitors are typically evaluated on bespoke, narrowly scoped datasets, leaving open questions about their accuracy, generalizability, and trade-offs.

Auricle is a dataset comprising synchronous noninvasive cardiovascular signals collected from 42 healthy subjects and 28 people with lived experience of heart failure. The dataset includes ECG, dual-site PPG, accelerometer-based SCG, digital stethoscope audio, binaural in-ear cardiac audio, continuous blood pressure, and respiratory effort. Most participants in the healthy cohort completed two 30-minute sessions on different days, while heart failure cohort participants completed one session. During each session, participants performed at least three types of activities intended to alter their cardiovascular function. Auricle enables researchers to evaluate new wearable cardiovascular monitoring methods using a standardized benchmark with significant intra- and inter-subject variability. Starter code for exploring and processing this dataset is available on GitHub [26, 27].

Background

Wearable technologies like smartwatches [1–3], smart rings [4–6], and wireless earbuds [7–9] have been proposed for passive, noninvasive cardiovascular health monitoring. These devices typically measure phenomena associated with the pulse wave using inertial, acoustic, or optical signals, or a combination thereof. The timing of pulse arrival at a single location can be used to estimate heart rate and heart rate variability [7], while the timing differences across multiple locations can be used to estimate pulse wave velocity as a correlate of blood pressure [10]. Analyzing the shape of these waveforms also yields a class of techniques known as pulse wave analysis [10, 11], which enables the estimation of more sophisticated cardiovascular parameters (e.g., cardiac output, stroke volume, systemic vascular resistance).

Evaluations of wearable cardiovascular monitors often use bespoke, narrowly scoped datasets. This not only makes it difficult to compare results, but can also constrain the development of innovative multimodal solutions. Auricle is designed to address these gaps with the following key features:

1. Broad Signal Coverage: Auricle comprises eleven synchronous cardiovascular waveforms, enabling both unimodal and multimodal explorations. The most distinctive signals in our dataset are binaural in-ear cardiac audio recordings using modified commercial ANC earbuds. In-ear cardiac audio is produced by a combination of proximal arterial expansion and distal sound transduction [12]. Auricle also includes respiratory effort, providing a reference for respiratory activity. In addition to physiological signals, the dataset includes several derived cardiovascular parameters such as heart rate, blood pressure, and cardiac output.
2. High Sample Rate: All physiological signals in Auricle are sampled at 1 kHz, far exceeding the typical sampling rates of consumer-grade devices. For blood pressure applications specifically, Auricle provides both periodic arm-cuff measurements and continuous finger-cuff measurements.
3. Diverse Participant Cohorts: Studies focusing on basic cardiovascular parameters like heart rate and blood pressure are often conducted with young, healthy subjects. Auricle not only covers a wide range of ages (19–79 years) but also includes people with lived experience of heart failure.
4. Intra-Subject Variability: To support data-driven model calibration and rigorous evaluation, subjects in Auricle's healthy cohort were invited to complete sessions on two separate days. Participants from both cohorts performed multiple activities to alter their cardiovascular function (exercise, breath-holding, Valsalva maneuver).

Auricle provides a valuable resource for biomedical engineers and ubiquitous computing researchers to benchmark existing and novel cardiovascular monitors. Moreover, machine learning researchers can leverage Auricle to prototype cutting-edge machine learning models for processing multimodal time-series data.

Methods

Auricle was created using a three-step process comprising participant recruitment, multimodal data collection, and post-processing.

Recruitment.

Auricle comprises data collected from two cohorts in Toronto, Ontario:

1. Healthy Cohort: 42 subjects aged 19-55 (mean: 27.4 ± 9.5 years) were recruited from the general public to contribute data at the University of Toronto. They were asked to complete two one-hour sessions on different days, but only 37 (88%) returned for the second session.
2. Heart Failure Cohort: 28 subjects aged 21-79 (mean: 52.2 ± 15.2 years) were recruited during routine medical appointments at the University Health Network's Ted Rogers Centre for Heart Research. Since many of these individuals traveled from outside the city for these appointments, subjects in this cohort completed only one session.

Data Collection.

After completing a brief demographics form, subjects' height and weight were measured. Subjects in the healthy cohort were also photographed in front of a gridded backdrop, enabling the measurement of body dimensions that may be useful for calibration (e.g., arm length, torso width). Cardiovascular data were acquired using a BIOPAC data acquisition system comprising an MP160 module [28] and 100 Series sensor-specific amplifiers. The following sensors were used to capture data at 10 kHz:

1. Electrocardiography: A set of 3M 2560 electrodes [29] was placed in a 3-lead configuration.
2. Phonocardiography: A Thinklabs One [30] digital stethoscope was attached to the chest using flexible straps and a 3D-printed fixture.
3. Seismocardiography: A BIOPAC 3-axis accelerometer [31] was affixed to the chest with double-sided tape.
4. Photoplethysmography (Finger): A Nonin PPG sensor [32] was clipped onto the left index finger in transmissive mode.
5. Photoplethysmography (Ear): A Nonin PPG sensor [32] was clipped onto the left earlobe in transmissive mode.
6. In-Ear Cardiac Audio: Modified Sony WF-1000XM3 earbuds [33] were placed in each ear. The earbuds' internal electronics were removed, and the in-ear condenser microphones were wired out for data acquisition.
7. Continuous Arterial Pressure: A continuous noninvasive arterial pressure (CNAP) monitor supported by BIOPAC [34] was worn on the right hand with the finger cuffs on the index and middle fingers. The monitor measures the arterial pressure waveform continuously, allowing it to report continuous systolic and diastolic blood pressure readings. Midway through the study, a software update enabled the monitor to also report cardiac output, stroke volume, and systemic vascular resistance; consequently, this information is available for a subset of subjects.
8. Respiratory Effort: A BIOPAC respiration belt [35] was wrapped around the torso to track breathing.

Each data collection session lasted approximately one hour. The first 30 minutes were typically consumed in setup and calibration, and the remaining 30 minutes were used for data collection. Once the sensors were properly affixed, subjects alternated between periods of rest and cardiovascular stress activities designed to induce variation in their waveforms. The healthy cohort performed four activities: a breath hold (4 repetitions), light exercise (2 repetitions), the cold pressor test (1 repetition), and the Valsalva maneuver (4 repetitions). The heart failure cohort did not complete the cold pressor test in order to reduce physical risk. The order of these activities was assigned to each participant using a Latin square in order to mitigate ordering effects.

A researcher continuously monitored physiological signals for quality assurance, but the nature of the protocol and the abundance of sensors occasionally resulted in temporary signal degradation. This was particularly true of the earbud audio. One earbud was prioritized in some of these cases to ensure that at least one ear recorded high-quality in-ear cardiac audio.

Post-Processing.

To protect participant privacy, all collected data were de-identified in accordance with HIPAA standards. Subject data are associated with an alphanumeric identifier (e.g., P1, P2) rather than identifiable information such as names or email addresses. Subjects were informed that the microphones in the earbuds and stethoscopes could record their voices, so they were encouraged to minimize speech. However, there were instances when either the subjects or the researchers needed to make adjustments to improve the quality of data. We made every effort to redact speech using a combination of signal processing and manual inspection. The audio signals were first processed using Silero VAD [13] to identify speech segments. The identified segments were extended by 2 seconds (before and after) to account for marginal or misaligned detections, after which they were redacted from the recordings. A researcher then scanned the audio files in a digital audio editing application to identify and redact any lingering speech segments.

After redacting speech, all continuous signals were downsampled to a common sampling rate of 1 kHz. The CNAP data include a masking array that identifies regions where data are known to be unreliable due to automatic calibration events or other obvious signal artifacts. The first few minutes, during which the CNAP device had not yet been calibrated, are marked in this mask. Other sensors were active and valid during this period.

Data Description

The dataset is organized such that physiological signals, stressor activity labels, and CNAP system annotations collected during a single session are stored in the WFDB format. All files are stored in folders named as follows: <root>/data/{cohort_id}_{subject_num}_{session_num}. The three bracketed placeholders are defined as:

1. cohort_id: HC for the healthy cohort, HFC for the heart failure cohort.
2. subject_num: The numerical identifier assigned to the subject (1–42 for the healthy cohort, 101–149 for the heart failure cohort).
3. session_num: A numerical identifier assigned to each subject-specific session.

Each folder includes the following files:

• Data ({cohort_id}_{subject_num}_{session_num}.dat): Time-series physiological signals corresponding to each sensor, hemodynamic metrics derived from the CNAP (e.g., systolic blood pressure, cardiac output), and annotation (e.g., stressor activities) bundled into a WFDB record.
• WFDB Header File ({cohort_id}_{subject_num}_{session_num}.hea): Metadata related to the WFDB record.

There are also three dataset-level files in Auricle's root directory, including:

• demographics.csv: A file containing demographic data comprising sex, height, weight, and age, as well as clinical measurements for the heart failure cohort.
• activity_labels.csv: A list of numeric values associated with cardiovascular stressor activities. These values are referenced in each session's "Activity-Annotations" channel.
• sensor_coordinates.csv: A list of spatial coordinates (in centimeters) describing the location of anatomical landmarks and sensors extracted from images of healthy cohort participants. These coordinates are expressed in terms of distance from the participant's centerline (x) and height (y).

Usage Notes

Auricle can be leveraged for research in several topic areas, including wearable sensing, machine learning, digital health, and biomedical engineering. Prospective areas of exploration include the following:

• Comparing proposed wearable systems
  Proposed wearable cardiovascular monitors [1–9, 14–17] have often been evaluated using closed datasets collected by the authors. Auricle allows researchers to compare them using a common dataset that contains greater inter- and intra-subject variability than is often seen in previous evaluations.
• Regressing understudied cardiovascular health metrics
  Prior work has often focused on estimating heart rate [7–9], heart rate variability [7–9], and blood pressure [2, 3, 6]. Auricle can be utilized to investigate underexplored cardiovascular metrics such as cardiac output, stroke volume, and systemic vascular resistance.
• Calibrating models
  Data-driven cardiovascular models often benefit from personalization, whether through fixed parameters (e.g., age, height, weight) or through subject-specific model fine-tuning. Auricle includes data that could be used to examine and evaluate these techniques, including but not limited to relative sensor distances and data collected across multiple sessions.
• Exploring in-ear cardiac audio
  Auricle includes binaural in-ear cardiac audio, a nascent modality for cardiac monitoring. This signal has been used primarily for heart rate monitoring [7–9, 14, 18] and pulse wave reconstruction [19], leaving the door open for other applications.
• Harnessing multimodality
  The signals in our dataset hold complementary information about cardiovascular function. Additionally, the information captured by a single sensor can be extended using deep learning models trained for signal reconstruction (e.g., PPG from ECG [20], PCG from in-ear cardiac audio [21], pulse wave from in-ear cardiac audio [19]). Auricle enables researchers to train multimodal and cross-modal models spanning diverse sensor combinations for comprehensive cardiovascular monitoring.

The dataset has notable limitations, including:

• Cross-Cohort Demographics
  Our two cohorts do not have the same demographic composition. The heart failure cohort is typically older than the healthy cohort and is predominantly male. These demographic biases limit what claims can be made about heart failure classification algorithms trained using Auricle. While a subset of each cohort can be roughly age-matched, this subset is much smaller than the overall dataset. However, Auricle remains more diverse than many datasets used to demonstrate the feasibility of cutting-edge wearables.
• Limitations of CNAP Devices
  CNAP monitors are not a gold-standard cardiovascular reference and can be sensitive to poor calibration. More reliable methods such as echocardiography and intra-arterial catheterization are costly, carry risk, and are often impractical for use during cardiovascular stress activities. Their use would have severely limited our ability to collect multimodal data during these activities. While not the gold standard, CNAP monitors produce outputs that are highly correlated with intra-arterial blood pressure [22, 23] and cardiac output [24] across clinical settings, and they avoid the discrete-measurement limitation associated with some traditional alternatives (e.g., white-coat hypertension [25]).
• Waveform Quality
  Waveform quality in Auricle is varied. A research team member monitored recorded physiological data in real time in order to prevent poor recordings by adjusting dislodged sensors, but this mitigation was imperfect. Consequently, some sensors experienced intermittent signal degradation. Additionally, motion artifacts were unavoidable because of the sensors' sensitivity to body movement.

An interactive interface for dataset exploration [26] and a processing Python library [27] are available on GitHub. The library, PulseKit, demonstrates how to load and visualize signals while providing a pipeline to segment cardiac cycles, assess quality via morphological features, and align them across modalities.

Release Notes

Version 1.0.0

Initial release

Ethics

The study protocol for data collection and public release was reviewed and approved by the Research Ethics Boards at the University of Toronto (Protocol #41568) and the University Health Network (#24-5431), which granted approval for the study and the sharing of de-identified data.

Acknowledgements

The authors gratefully acknowledge financial support from TRANSFORM HF and its Digital Hub at the University Health Network.

Conflicts of Interest

We report no conflicts of interest.

References

1. Carek, A. M., Conant, J., Joshi, A., Kang, H., & Inan, O. T. (2017). SeismoWatch: wearable cuffless blood pressure monitoring using pulse transit time. Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies, 1(3), 1-16.
2. Cao, Y., Chen, H., Li, F., & Wang, Y. (2021, October). Crisp-BP: Continuous wrist PPG-based blood pressure measurement. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking (pp. 378-391).
3. Li, X., Hussein, R., Zhu, G., Sui, X., Li, H., Yang, X., ... & Li, Y. (2024, July). Continuous blood pressure monitoring and hypertension risk screening using smart watch. In 2024 46th annual international conference of the IEEE engineering in medicine and biology society (EMBC) (pp. 1-6). IEEE.
4. Montenegro, M., Aliverti, A., & Angelucci, A. (2026). An adjustable smart ring to monitor pulse rate and peripheral blood oxygen saturation. Annals of Biomedical Engineering, 54(3), 885-897.
5. Sel, K., Osman, D., Huerta, N., Edgar, A., Pettigrew, R. I., & Jafari, R. (2023). Continuous cuffless blood pressure monitoring with a wearable ring bioimpedance device. npj Digital Medicine, 6(1), 59.
6. Zhu, G., Zhao, D., Song, Y., Wang, Z., Zhou, A., Ma, H., ... & Tang, Y. (2025). RingBP: Towards continuous, comfortable, and generalized blood pressure monitoring using a smart ring. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 9(2), 1-24.
7. Butkow, K. J., Dang, T., Ferlini, A., Ma, D., & Mascolo, C. (2023, March). heart: Motion-resilient heart rate monitoring with in-ear microphones. In 2023 IEEE International Conference on Pervasive Computing and Communications (PerCom) (pp. 200-209). IEEE.
8. Fan, X., Pearl, D., Howard, R., Shangguan, L., & Thormundsson, T. (2023, October). Apg: Audioplethysmography for cardiac monitoring in hearables. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking (pp. 1-15).
9. Gilliam III, F. R., Ciesielski, R., Shahinyan, K., Shakya, P., Cunsolo, J., Panchal, J. M., ... & Bhavnani, S. P. (2022). In-ear infrasonic hemodynography with a digital health device for cardiovascular monitoring using the human audiome. NPJ Digital Medicine, 5(1), 189.
10. Mukkamala, R., Stergiou, G. S., & Avolio, A. P. (2022). Cuffless blood pressure measurement. Annual Review of Biomedical Engineering, 24(1), 203-230.
11. Wang, L., Pickwell-MacPherson, E., Liang, Y. P., & Zhang, Y. T. (2009, September). Noninvasive cardiac output estimation using a novel photoplethysmogram index. In 2009 annual international conference of the IEEE engineering in medicine and biology society (pp. 1746-1749). IEEE.
12. Christofferson, K., Bhalla, S., Cafazzo, J., & Mariakakis, A. (2024, October). On the production and measurement of cardiac sounds in the ear canal. In Companion of the 2024 on ACM International Joint Conference on Pervasive and Ubiquitous Computing (pp. 685-690).
13. Team, S. (2021, March). Silero vad: pre-trained enterprise-grade voice activity detector (vad), number detector and language classifier.
14. Truong, H., Montanari, A., & Kawsar, F. (2022, May). Non-invasive blood pressure monitoring with multi-modal in-ear sensing. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 6-10). IEEE.
15. Kusche, R., Klimach, P., Malhotra, A., Kaufmann, S., & Ryschka, M. (2015). An in-ear pulse wave velocity measurement system using heart sounds as time reference. Current Directions in Biomedical Engineering, 1(1), 366-370.
16. Zhao, Z., Li, F., Xie, Y., Xie, H., Zhang, K., Zhang, L., & Wang, Y. (2024, May). HearBP: Hear your blood pressure via in-ear acoustic sensing based on heart sounds. In IEEE INFOCOM 2024-IEEE Conference on Computer Communications (pp. 991-1000). IEEE.
17. Ciesielski, R., Wong, C. S., Panchal, J., Grela, J., Dimalanta, C., Walton, A., ... & Barnacka, A. (2025). Non-invasive blood pressure monitoring with in-ear infrasonic hemodynography for preventative cardiovascular care. npj Cardiovascular Health, 2(1), 39.
18. Nirjon, S., Dickerson, R. F., Li, Q., Asare, P., Stankovic, J. A., Hong, D., ... & Zhao, F. (2012, November). Musicalheart: A hearty way of listening to music. In Proceedings of the 10th ACM conference on embedded network sensor systems (pp. 43-56).
19. Christofferson, K., Lin, M., Cafazzo, J. A., & Mariakakis, A. (2025). ArtEARial: Arterial Pressure Waveform Reconstruction Using Earbud Audio. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 9(4), 1-24.
20. Zhu, Q., Tian, X., Wong, C. W., & Wu, M. (2021). Learning your heart actions from pulse: ECG waveform reconstruction from PPG. IEEE Internet of Things Journal, 8(23), 16734-16748.
21. Cao, A., Christofferson, K., Ruth, P., Rabbani, N., Shi, Y., Mariakakis, A., ... & Patel, S. (2024, July). Earsteth: Cardiac auscultation audio reconstruction using earbuds. In 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) (pp. 1-4). IEEE.
22. Jeleazcov, C., Krajinovic, L., Münster, T., Birkholz, T., Fried, R., Schüttler, J., & Fechner, J. (2010). Precision and accuracy of a new device (CNAP™) for continuous non-invasive arterial pressure monitoring: assessment during general anaesthesia. British journal of anaesthesia, 105(3), 264-272.
23. Fortin, J., Wellisch, A., & Maier, K. (2013). CNAP-evolution of continuous non-invasive arterial blood pressure monitoring. Biomedical Engineering/Biomedizinische Technik, 58, 1.
24. Wagner, J. Y., Grond, J., Fortin, J., Negulescu, I., Schöfthaler, M., & Saugel, B. (2016). Continuous noninvasive cardiac output determination using the CNAP system: evaluation of a cardiac output algorithm for the analysis of volume clamp method-derived pulse contour. Journal of clinical monitoring and computing, 30(4), 487-493.
25. Briasoulis, A., Androulakis, E., Palla, M., Papageorgiou, N., & Tousoulis, D. (2016). White-coat hypertension and cardiovascular events: a meta-analysis. Journal of hypertension, 34(4), 593-599.
26. Christofferson, K. (2026). Auricle Data Visualizer [Computer software]. GitHub. https://github.com/chai-toronto/Auricle-Data-Visualizer
27. Kleinknecht, T., Christofferson, K., & Mariakakis, A. (2026). PulseKit [Computer software]. GitHub. https://github.com/chai-toronto/pulsekit
28. https://www.biopac.com/product/mp160-system-with-acqknowledge/
29. https://multimedia.3m.com/mws/media/1415893O/3m-red-dot-electrodes-ecg-electrodes-product-comparison-chart.pdf
30. https://store.thinklabs.com/products/thinklabs-one-digital-stethoscope
31. https://www.biopac.com/product/accelerometer-2g/
32. https://www.nonin.com/
33. https://www.sony.com/ke/electronics/truly-wireless/wf-1000xm3
34. https://www.biopac.com/product/noninvasive-blood-pressure-amp-with-hemodynamics/
35. https://www.biopac.com/product/respiratory-effort-trans-tp/