pith. sign in

arxiv: 1907.07496 · v1 · pith:R7QQQJDYnew · submitted 2019-07-17 · 💻 cs.LG · cs.HC· stat.ML

Improving Heart Rate Variability Measurements from Consumer Smartwatches with Machine Learning

Martin Maritsch , Caterina B\'erub\'e , Mathias Kraus , Vera Lehmann , Thomas Z\"uger , Stefan Feuerriegel , Tobias Kowatsch , Felix Wortmann This is my paper

Pith reviewed 2026-05-24 20:21 UTC · model grok-4.3

classification 💻 cs.LG cs.HCstat.ML
keywords heart rate variabilitysmartwatchesmachine learningaccelerometer datawearable sensorserror correctionmovement artifact
0
0 comments X

The pith

Smartwatch HRV errors correlate with wearer movement and can be reduced by machine learning on accelerometer data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper investigates heart rate variability measurements taken by consumer smartwatches and identifies a systematic error tied to the wearer's movement. It demonstrates a statistically significant correlation between this error and movement intensity. The work then shows that the error decreases when additional on-device sensor streams, such as accelerometer readings, are fed into a neural model. This correction approach is presented as a way to make continuous HRV tracking more reliable for everyday assessment of physical and mental health without requiring new hardware.

Core claim

The central claim is that error in smartwatch HRV readings is not random but systematically linked to wearer movement, and that this bias can be learned and subtracted by bringing accelerometer and related sensor data into a neural learning model.

What carries the argument

Neural learning applied to the combination of raw HRV signals and simultaneous accelerometer data to predict and remove movement-dependent measurement bias.

Load-bearing premise

The error observed in HRV readings is a repeatable, movement-dependent bias that additional device sensors can capture and a model can learn to subtract.

What would settle it

A controlled experiment that measures the same heart signal simultaneously with a medical-grade device and a smartwatch across varying movement levels, then checks whether the proposed model still leaves a statistically significant residual error after correction.

Figures

Figures reproduced from arXiv: 1907.07496 by Caterina B\'erub\'e, Felix Wortmann, Martin Maritsch, Mathias Kraus, Stefan Feuerriegel, Thomas Z\"uger, Tobias Kowatsch, Vera Lehmann.

Figure 1
Figure 1. Figure 1: Samples of the RMSSD as calculated from data of the heart rate monitor (black, reference value) and the consumer [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
read the original abstract

The reactions of the human body to physical exercise, psychophysiological stress and heart diseases are reflected in heart rate variability (HRV). Thus, continuous monitoring of HRV can contribute to determining and predicting issues in well-being and mental health. HRV can be measured in everyday life by consumer wearable devices such as smartwatches which are easily accessible and affordable. However, they are arguably accurate due to the stability of the sensor. We hypothesize a systematic error which is related to the wearer movement. Our evidence builds upon explanatory and predictive modeling: we find a statistically significant correlation between error in HRV measurements and the wearer movement. We show that this error can be minimized by bringing into context additional available sensor information, such as accelerometer data. This work demonstrates our research-in-progress on how neural learning can minimize the error of such smartwatch HRV measurements.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript claims that HRV measurements from consumer smartwatches contain a systematic, movement-related error that can be detected via statistically significant correlation with wearer movement and then reduced by incorporating additional on-device sensor data (e.g., accelerometer) into a neural model. The work is presented as research-in-progress demonstrating that explanatory and predictive modeling can minimize this error.

Significance. If the central claim were substantiated with an independent reference standard and a properly held-out evaluation, the result would be relevant to the growing literature on artifact correction in wearable PPG signals. However, the absence of any description of the ground-truth HRV reference, dataset, model architecture, or validation protocol means the reported improvement cannot be assessed for independence from physiology or from training-set fit.

major comments (2)
  1. [Abstract] Abstract: The manuscript asserts a 'statistically significant correlation between error in HRV measurements and the wearer movement' and that 'this error can be minimized' by ML, yet supplies no description of the reference device, protocol, or ground-truth HRV used to define the error term. Without this, any observed correlation is consistent with both motion artifact and genuine autonomic changes during movement; only the former is correctable by the proposed approach.
  2. [Abstract] Abstract / Methods (missing): No model specification, training procedure, baseline comparison, error bars, dataset size, or cross-validation scheme is provided. The central claim of successful error reduction therefore cannot be evaluated and the reported improvement may simply reflect training-set fit rather than an independent test.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their review. The manuscript is explicitly a short research-in-progress report, which explains the absence of full methodological details. We agree that these omissions prevent proper evaluation of the claims and will expand the work accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The manuscript asserts a 'statistically significant correlation between error in HRV measurements and the wearer movement' and that 'this error can be minimized' by ML, yet supplies no description of the reference device, protocol, or ground-truth HRV used to define the error term. Without this, any observed correlation is consistent with both motion artifact and genuine autonomic changes during movement; only the former is correctable by the proposed approach.

    Authors: We agree that the current abstract provides no description of the reference device, protocol, or ground-truth definition. As a brief research-in-progress note, the text focuses on the hypothesis rather than experimental details. The error term is computed as the difference between smartwatch-derived HRV and a reference measurement, but without the requested information it is impossible to exclude physiological confounds. We will revise the manuscript to specify the reference (clinical ECG), protocol (rest vs. movement tasks), and exact HRV metric used. revision: yes

  2. Referee: [Abstract] Abstract / Methods (missing): No model specification, training procedure, baseline comparison, error bars, dataset size, or cross-validation scheme is provided. The central claim of successful error reduction therefore cannot be evaluated and the reported improvement may simply reflect training-set fit rather than an independent test.

    Authors: We agree that the manuscript supplies none of the listed methodological elements. This omission is a direct consequence of the short research-in-progress format. The neural model uses accelerometer features as additional inputs to a regression network, but without architecture, dataset size, validation scheme, or baselines the improvement cannot be assessed. We will add these specifications, including subject-wise cross-validation and held-out performance metrics, in the next version. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation remains self-contained.

full rationale

The abstract and description present a hypothesis of movement-related systematic error in HRV, a reported correlation, and an ML correction using accelerometer data. No equations, fitted parameters renamed as predictions, self-citations, or uniqueness claims are supplied that would reduce any result to its inputs by construction. The modeling step is described at a high level without evidence that the reported improvement collapses to a training fit on the same pairs or any other enumerated circular pattern. This is the normal case of an independent empirical claim whose validity can be assessed externally.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the untested premise that movement produces a learnable, additive bias in the optical HRV signal that is independent of other physiological sources of variability. No free parameters, axioms, or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption Error in optical HRV is systematically related to wearer movement and can be isolated from other sources of variability.
    Stated as the working hypothesis that motivates the modeling approach.

pith-pipeline@v0.9.0 · 5703 in / 1268 out tokens · 20887 ms · 2026-05-24T20:21:22.389454+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 1 internal anchor

  1. [1]

    Hyun Jae Baek, Chul-Ho Cho, Jaegeol Cho, and Jong-Min Woo. 2015. Reliability of Ultra-Short-Term Analysis as a Surrogate of Standard 5-Min Analysis of Heart Rate Variability. Telemedicine and e-Health 21, 5 (2015), 404–414. https://doi.org/10.1089/tmj.2014.0104

    work page doi:10.1089/tmj.2014.0104 2015

  2. [2]

    Hejlesen, Lise Tarnow, and Jesper Fleischer

    Simon Lebech Cichosz, Jan Frystyk, Ole K. Hejlesen, Lise Tarnow, and Jesper Fleischer. 2014. A novel algorithm for prediction and detection of hypoglycemia based on continuous glucose monitoring and heart rate variability in patients with type 1 diabetes. Journal of Diabetes Science and Technology 8, 4 (2014), 731–737. https://doi.org/10.1177/ 1932296814528838

    work page 2014

  3. [3]

    Philip E. Cryer. 2004. Diverse Causes of Hypoglycemia-Associated Autonomic Failure in Diabetes. New England Journal of Medicine 350, 22 (2004), 2272–2279. https://doi.org/10.1056/NEJMra031354

    work page doi:10.1056/nejmra031354 2004

  4. [4]

    Dimitriev and E.V

    D.A. Dimitriev and E.V. Saperova. 2015. Heart rate variability as a measure of autonomic regulation of cardiac activity for assessing mental stress. Autonomic Neuroscience 192, 3 (2015), 80. https://doi. org/10.1016/j.autneu.2015.07.086

    work page doi:10.1016/j.autneu.2015.07.086 2015

  5. [5]

    Erin E Dooley, Natalie M Golaszewski, and John B Bartholomew. 2017. Estimating Accuracy at Exercise Intensities: A Comparative Study of Self-Monitoring Heart Rate and Physical Activity Wearable Devices. JMIR mHealth and uHealth 5, 3 (2017), e34. https://doi.org/10.2196/ mhealth.7043

    work page 2017

  6. [6]

    Fatema El-Amrawy and Mohamed Ismail Nounou. 2015. Are currently available wearable devices for activity tracking and heart rate monitor- ing accurate, precise, and medically beneficial? Healthcare Informatics Research 21, 4 (2015), 315–320. https://doi.org/10.4258/hir.2015.21.4. 315

    work page doi:10.4258/hir.2015.21.4 2015

  7. [7]

    Esco and Andrew A

    Michael R. Esco and Andrew A. Flatt. 2014. Ultra-short-term heart rate variability indexes at rest and post-exercise in athletes: Evaluating the agreement with accepted recommendations. Journal of Sports Science and Medicine 13, 3 (2014), 535–541

    work page 2014

  8. [8]

    International Diabetes Federation. 2017. IDF Diabetes Atlas (8 ed.). International Diabetes Federation. http://diabetesatlas.org

    work page 2017

  9. [9]

    Z. Ge, P. W.C. Prasad, N. Costadopoulos, Abeer Alsadoon, A. K. Singh, and A. Elchouemi. 2016. Evaluating the accuracy of wearable heart rate monitors. In Proceedings - 2016 International Conference on Advances Improving HRV Measurements from Smartwatches with Machine Learning in Computing, Communication and Automation (Fall), ICACCA 2016 . IEEE, 1–6. http...

    work page doi:10.1109/icaccaf.2016.7748986 2016

  10. [10]

    Marc Gillinov, Muhammad Etiwy, Stephen Gillinov, Robert Wang, Gordon Blackburn, Dermot Phelan, Penny Houghtaling, Hoda Javadikasgari, and Milind Y

    A. Marc Gillinov, Muhammad Etiwy, Stephen Gillinov, Robert Wang, Gordon Blackburn, Dermot Phelan, Penny Houghtaling, Hoda Javadikasgari, and Milind Y. Desai. 2017. Variable Accuracy of Commer- cially Available Wearable Heart Rate Monitors.Journal of the American College of Cardiology 69, 11 (2017), 336. https://doi.org/10.1016/s0735- 1097(17)33725-7

    work page doi:10.1016/s0735- 2017

  11. [11]

    Goldberger, Sridevi Challapalli, Roderick Tung, Michele A

    Jeffrey J. Goldberger, Sridevi Challapalli, Roderick Tung, Michele A. Parker, and Alan H. Kadish. 2001. Relationship of heart rate variability to parasympathetic effect. Circulation 103, 15 (2001), 1977–1983. https: //doi.org/10.1161/01.CIR.103.15.1977

    work page doi:10.1161/01.cir.103.15.1977 2001

  12. [12]

    André Henriksen, Martin Haugen Mikalsen, Ashenafi Zebene Woldare- gay, Miroslav Muzny, Gunnar Hartvigsen, Laila Arnesdatter Hopstock, and Sameline Grimsgaard. 2018. Using fitness trackers and smart- watches to measure physical activity in research: Analysis of consumer wrist-worn wearables. Journal of Medical Internet Research 20, 3 (2018), e110. https://...

    work page doi:10.2196/jmir.9157 2018

  13. [13]

    Mordor Intelligence. 2018. Smart Watch Market - Growth, Trends, and Forecast (2019 - 2024). https://www.mordorintelligence.com/industry- reports/global-smart-watches-market-industry

    work page 2018

  14. [14]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Ima- geNet Classification with Deep Convolutional Neural Networks. In Advances in neural information processing systems . 1097–1105

    work page 2012

  15. [15]

    Marek Malik. 1996. Heart rate variability. Standards of measurement, physiological interpretation, and clinical use. Task Force of the Euro- pean Society of Cardiology and the North American Society of Pacing and Electrophysiology. European Heart Journal 17, 3 (1996), 354–81. http://www.ncbi.nlm.nih.gov/pubmed/8737210

    work page arXiv 1996

  16. [16]

    Loretto Munoz, Arie Van Roon, Harriëtte Riese, Chris Thio, Emma Oostenbroek, Iris Westrik, Eco J.C

    M. Loretto Munoz, Arie Van Roon, Harriëtte Riese, Chris Thio, Emma Oostenbroek, Iris Westrik, Eco J.C. De Geus, Ron Gansevoort, Joop Lefrandt, Ilja M. Nolte, and Harold Snieder. 2015. Validity of (Ultra- )Short recordings for heart rate variability measurements. PLoS ONE 10, 9 (2015), e0138921. https://doi.org/10.1371/journal.pone.0138921

    work page doi:10.1371/journal.pone.0138921 2015

  17. [17]

    Udi Nussinovitch, Keren Politi Elishkevitz, Keren Katz, Moshe Nussi- novitch, Shlomo Segev, Benjamin Volovitz, and Naomi Nussinovitch

    work page

  18. [18]

    Annals of Noninvasive Electrocardiology 16, 2 (2011), 117–122

    Reliability of ultra-short ECG indices for heart rate variabil- ity. Annals of Noninvasive Electrocardiology 16, 2 (2011), 117–122. https://doi.org/10.1111/j.1542-474X.2011.00417.x

    work page doi:10.1111/j.1542-474x.2011.00417.x 2011

  19. [19]

    Jakub Parak and Ilkka Korhonen. 2013. Accuracy of Firstbeat Body- guard 2 beat-to-beat heart rate monitor. (Whitepaper) (2013), 6–

    work page 2013

  20. [20]

    https://assets.firstbeat.com/firstbeat/uploads/2015/10/white_paper_ bodyguard2_final.pdf

    work page 2015

  21. [21]

    Pathirana, and Aruna Senevi- ratne

    Dung Phan, Lee Yee Siong, Pubudu N. Pathirana, and Aruna Senevi- ratne. 2015. Smartwatch: Performance evaluation for long-term heart rate monitoring. In 4th International Symposium on Bioelectronics and Bioinformatics, ISBB 2015. 144–147. https://doi.org/10.1109/ISBB.2015. 7344944

    work page doi:10.1109/isbb.2015 2015

  22. [22]

    Dung Phan, Lee Yee Siong, Pubudu N Pathirana, and Aruna Senevi- ratne. 2015. Smartwatch: Performance evaluation for long-term heart rate monitoring - IEEE Conference Publication. In Ieeexplore.Ieee.Org. IEEE, 144–147. https://ieeexplore.ieee.org/document/7344944/

    work page arXiv 2015

  23. [23]

    Lizawati Salahuddin, Jaegeol Cho, Myeong Gi Jeong, and Desok Kim

    work page

  24. [24]

    In Annual International Conference of the IEEE Engineering in Medicine and Biology - Proceedings

    Ultra short term analysis of heart rate variability for monitoring mental stress in mobile settings. In Annual International Conference of the IEEE Engineering in Medicine and Biology - Proceedings . IEEE, 4656–4659. https://doi.org/10.1109/IEMBS.2007.4353378

    work page doi:10.1109/iembs.2007.4353378 2007

  25. [25]

    Wilhelm, Peter R

    Hartmut Schächinger, Johannes Port, Stuart Brody, Lilly Linder, Frank H. Wilhelm, Peter R. Huber, Daniel Cox, and Ulrich Keller. 2004. Increased high-frequency heart rate variability during insulin-induced hypoglycaemia in healthy humans. Clinical Science 106, 6 (2004), 583–

    work page 2004

  26. [26]

    https://doi.org/10.1042/cs20030337

    work page doi:10.1042/cs20030337

  27. [27]

    Thong, K

    T. Thong, K. Li, J. McNames, M. Aboy, and B. Goldstein. 2004. Accuracy of ultra-short heart rate variability measures. InProceedings of the 25th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (IEEE Cat. No. 03CH37439) , Vol. 3. IEEE, 2424–2427. https://doi.org/10.1109/iembs.2003.1280405

    work page doi:10.1109/iembs.2003.1280405 2004

  28. [28]

    Martin Zihlmann, Dmytro Perekrestenko, and Michael Tschannen

    work page

  29. [29]

    Convolutional Recurrent Neural Networks for Electrocardiogram Classification

    Convolutional Recurrent Neural Networks for Electrocardiogram Classification. (2017), 1–4. arXiv:1710.06122 http://arxiv.org/abs/1710. 06122

    work page internal anchor Pith review Pith/arXiv arXiv 2017