pith. sign in

arxiv: 2605.00910 · v1 · submitted 2026-04-29 · 📡 eess.SP

Toward Real-Time Circadian Phase Estimation with Low Latency from Wearable Sensing Data

Mengzhu Xu , Nemanja Cabrilo , Merel van Gilst , Jean-Paul Linnartz This is my paper

Pith reviewed 2026-05-09 20:47 UTC · model grok-4.3

classification 📡 eess.SP
keywords circadian phase estimationwearable sensorsreal-time monitoringlight exposurephysical activitylow latencycosinor modelcore body temperature
0
0 comments X

The pith

A framework estimates circadian phase in real time from short wearable data windows with 1.19-hour error.

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

The paper tests whether circadian phase can be estimated instantly using only recent short windows of light exposure and physical activity from wearables, instead of waiting for a full day of recordings. It proposes a low-latency approach that predicts current phase from past observations and evaluates it against a cosinor-fitted core body temperature reference in a study of 14 free-living participants. Accuracy gets better as the history window grows but stops improving much after about 8 hours, and light-plus-activity data alone reaches a mean circular mean absolute error of 1.19 hours. This matters because current methods impose high latency and heavy data demands that prevent real-time use on edge devices for personalized health tracking. The work compares sensor choices, window lengths, and model types under participant-based cross-validation to guide practical deployment.

Core claim

The paper establishes that a low-latency framework can estimate instantaneous circadian phase from past observations of wearable data, with performance improving with window length but saturating near 8 hours of history; tree-based models plateau after 480 minutes while sequence models gain from longer contexts, and light exposure plus physical activity alone deliver a mean circular mean absolute error of 1.19 hours in a free-living study of 14 participants when referenced to cosinor-fitted core body temperature rhythm.

What carries the argument

The low-latency framework that predicts instantaneous circadian phase from historical windows of wearable sensor data, trained and tested via participant-based cross-validation and scored by circular mean absolute error against the cosinor reference.

If this is right

  • Accuracy improves with increasing window length but saturates at approximately 8 hours of history.
  • Tree-based models reach a performance plateau beyond 480 minutes whereas sequence-based models continue to benefit from longer temporal contexts.
  • The approach reduces data and computational burden, enabling real-time deployment on edge devices.
  • Light exposure and physical activity data alone achieve a mean circular mean absolute error of 1.19 hours.
  • The results supply practical guidance for choosing window lengths and model classes in wearable circadian monitoring.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Consumer wearables could integrate these estimates into sleep or activity apps to give immediate timing advice.
  • Adding heart-rate or skin-temperature channels might lower error further while keeping latency low.
  • The method could support timed interventions such as light therapy without the delay of full-cycle analysis.
  • Validation in larger and more varied populations would test whether the 8-hour saturation and 1.19-hour error generalize.

Load-bearing premise

The cosinor-fitted core body temperature rhythm serves as an accurate reference for true circadian phase in free-living conditions with wearable data from 14 participants.

What would settle it

Direct comparison of the estimates against gold-standard markers such as dim-light melatonin onset in a larger cohort, if it reveals average errors above 2 hours, would show the claimed accuracy does not hold.

Figures

Figures reproduced from arXiv: 2605.00910 by Jean-Paul Linnartz, Mengzhu Xu, Merel van Gilst, Nemanja Cabrilo.

Figure 1
Figure 1. Figure 1: Overview of the sensing setup used in the study. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall methodological workflow for low latency circadian phase estimation from wearable data. Signals from wearable and ingestible sensors are [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of window length and sensor modality on circadian phase [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example qualitative results for Participant#23 using the RF model [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Measured versus estimated circadian phase under daytime and [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
read the original abstract

Accurate estimation of the human circadian phase plays an important role in personalized health monitoring, but most existing wearable-based approaches operate retrospectively and require full circadian cycle recordings, leading to high estimation latency and substantial data and computational burden for real-time deployment on edge devices. In this study, we investigated whether circadian phase can be estimated in real time using only short historical windows of wearable data. We propose a low latency framework that estimates instantaneous circadian phase from past observations, with a cosinor-fitted core body temperature rhythm serving as the reference. Data from a free-living field study involving 14 participants were used to systematically evaluate the effects of sensor modality selection, historical window length, and model class under participant-based cross-validation. The results showed that estimation accuracy improves with increasing window length but saturates at approximately 8 hours of history. Tree-based models reached a performance plateau beyond 480 minutes, whereas sequence-based models continued to benefit from longer temporal contexts. When relying solely on light exposure and physical activity, the proposed approach achieved a mean circular mean absolute error (CMAE) of 1.19 h. These findings provide practical guidance for efficient and deployable real-time circadian phase monitoring using wearables.

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 / 2 minor

Summary. The paper proposes a low-latency framework for real-time circadian phase estimation from short historical windows of wearable data (light exposure and physical activity). Using participant-based cross-validation on free-living data from 14 subjects, it reports that estimation accuracy improves with window length but saturates near 8 hours, with tree-based models achieving a mean circular mean absolute error (CMAE) of 1.19 h when using only light and activity; a cosinor-fitted core body temperature (CBT) rhythm serves as the reference phase.

Significance. If the reference phase is reliable, the work offers practical guidance for deployable real-time circadian monitoring on edge devices by showing that 8 hours of history suffices and that light+activity alone can yield usable accuracy. The saturation result and modality comparison are useful for system design. However, the central performance claims rest on an unvalidated proxy reference, which weakens the significance for true circadian phase estimation.

major comments (2)
  1. [Abstract and Methods] Abstract and Methods (reference phase definition): The reference phase is obtained by cosinor fitting to CBT data, yet the manuscript provides no validation against gold-standard markers such as dim-light melatonin onset (DLMO) and does not discuss masking by activity, posture, meals, or sleep in free-living conditions. Because the reported CMAE of 1.19 h is measured exclusively against this proxy, the result may reflect consistency with a noisy or biased label rather than accuracy relative to true circadian phase.
  2. [Results] Results (model evaluation and saturation claim): The saturation of performance at approximately 8 hours (480 minutes) and the superiority of tree-based models are presented without sufficient detail on data preprocessing steps, feature definitions, hyperparameter selection, or statistical tests for differences across window lengths and modalities. These omissions make it impossible to verify that the 1.19 h CMAE and the plateau are robust rather than artifacts of the specific pipeline.
minor comments (2)
  1. [Abstract] The abstract states that sequence-based models continue to benefit from longer contexts while tree-based models plateau, but no supporting table or figure quantifies this divergence across the full range of window lengths tested.
  2. [Methods] Notation for circular mean absolute error (CMAE) should be defined explicitly with its formula, especially since the evaluation metric is central to all quantitative claims.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below, providing clarifications and indicating where revisions will be made.

read point-by-point responses

  1. Referee: [Abstract and Methods] Abstract and Methods (reference phase definition): The reference phase is obtained by cosinor fitting to CBT data, yet the manuscript provides no validation against gold-standard markers such as dim-light melatonin onset (DLMO) and does not discuss masking by activity, posture, meals, or sleep in free-living conditions. Because the reported CMAE of 1.19 h is measured exclusively against this proxy, the result may reflect consistency with a noisy or biased label rather than accuracy relative to true circadian phase.

    Authors: We agree that DLMO is widely regarded as the gold-standard marker for circadian phase. Our choice of cosinor-fitted CBT as the reference was driven by the practical requirements of a free-living wearable study, where continuous non-invasive temperature sensing is feasible while repeated DLMO sampling is not. Literature supports moderate-to-strong correlations between CBT-derived phase and DLMO (with typical offsets of 1-2 hours), making it a reasonable proxy for evaluating wearable-based estimation methods. In the revision we will (1) explicitly state in the abstract and methods that CBT serves as a proxy reference, (2) add a dedicated limitations paragraph discussing potential masking by activity, posture, meals, and sleep, and (3) cite supporting validation studies. We cannot, however, retroactively validate against DLMO because melatonin data were not collected. revision: partial

  2. Referee: [Results] Results (model evaluation and saturation claim): The saturation of performance at approximately 8 hours (480 minutes) and the superiority of tree-based models are presented without sufficient detail on data preprocessing steps, feature definitions, hyperparameter selection, or statistical tests for differences across window lengths and modalities. These omissions make it impossible to verify that the 1.19 h CMAE and the plateau are robust rather than artifacts of the specific pipeline.

    Authors: We acknowledge that the current manuscript lacks sufficient methodological transparency for full reproducibility. In the revised version we will expand the Methods section to detail: (a) all preprocessing steps (artifact removal, normalization, missing-value handling), (b) the complete feature set extracted from light and activity (e.g., mean, variance, percentiles, and temporal derivatives over sliding windows), (c) the hyperparameter search procedure (grid search with inner participant-based cross-validation), and (d) the statistical tests employed (repeated-measures ANOVA followed by Tukey post-hoc tests with correction for multiple comparisons) to evaluate differences across window lengths and sensor modalities. These additions will be placed in the main text or, if space-constrained, in supplementary material with clear pointers from the results. revision: yes

standing simulated objections not resolved
  • Direct validation of the CBT-derived reference phase against DLMO is not possible because melatonin onset data were not collected in this dataset.

Circularity Check

0 steps flagged

No significant circularity; supervised prediction against independent reference

full rationale

The paper trains ML models (tree-based and sequence-based) on short windows of light exposure and physical activity data to estimate instantaneous circadian phase. The target label is obtained by fitting a cosinor model to separate core body temperature recordings from the same participants. This is a standard supervised regression setup evaluated under participant-based cross-validation. No derivation step, equation, or self-citation reduces the reported CMAE of 1.19 h to a fitted parameter or input by construction. The reference phase is computed from temperature data that is not among the input features used for prediction. External benchmarks (cosinor fitting) are standard and independent of the wearable modalities being tested. The result is therefore self-contained and falsifiable against the held-out reference labels.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The claim rests on the assumption that cosinor fitting to core body temperature provides a valid phase reference and that ML models trained on 14 participants generalize; no new entities or many free parameters beyond standard model fitting are introduced.

free parameters (2)
  • historical window length = 8 hours
    Evaluated as saturating at approximately 8 hours (480 minutes) for tree-based models.
  • model class hyperparameters
    Tree-based and sequence models trained under cross-validation; specific values not detailed in abstract.
axioms (1)
  • domain assumption Cosinor model provides accurate reference for circadian phase from core body temperature
    Used as ground truth for evaluating wearable-based estimates in free-living data.

pith-pipeline@v0.9.0 · 5525 in / 1404 out tokens · 47446 ms · 2026-05-09T20:47:44.730804+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. [1]

    Circadian rhythm abnormal- ities,

    P. C. Zee, H. Attarian, and A. Videnovic, “Circadian rhythm abnormal- ities,”Continuum: Lifelong Learning in Neurology, vol. 19, no. 1, pp. 132–147, 2013

    work page 2013

  2. [2]

    C. A. Czeisler, R. E. Kronauer, J. S. Allan, J. F. Duffy, M. E. Jewett, E. N. Brown, and J. M. Ronda, “Bright light induction of strong (type

    work page

  3. [3]

    resetting of the human circadian pacemaker,”Science, vol. 244, no. 4910, pp. 1328–1333, 1989

    work page 1989

  4. [4]

    The national human activity pattern survey (nhaps): a resource for assessing exposure to environmental pollutants,

    N. E. Klepeis, W. C. Nelson, W. R. Ott, J. P. Robinson, A. M. Tsang, P. Switzer, J. V . Behar, S. C. Hern, and W. H. Engelmann, “The national human activity pattern survey (nhaps): a resource for assessing exposure to environmental pollutants,”Journal of Exposure Science and Environmental Epidemiology, vol. 11, no. 3, pp. 231–252, 2001

    work page 2001

  5. [5]

    Social jetlag: misalignment of biological and social time,

    M. Wittmann, J. Dinich, M. Merrow, and T. Roenneberg, “Social jetlag: misalignment of biological and social time,”Chronobiology International, vol. 23, no. 1–2, pp. 497–509, 2006

    work page 2006

  6. [6]

    Health implications of disrupted circadian rhythms and the potential for daylight as therapy,

    J. Brainard, M. Gobel, B. Scott, M. Koeppen, and T. Eckle, “Health implications of disrupted circadian rhythms and the potential for daylight as therapy,”Anesthesiology, vol. 122, no. 5, pp. 1170–1175, 2015

    work page 2015

  7. [7]

    Circadian misalignment and health,

    K. G. Baron and K. J. Reid, “Circadian misalignment and health,”Int Rev Psychiatry, vol. 26, no. 2, pp. 139–154, 2014

    work page 2014

  8. [8]

    Dim light melatonin onset (dlmo): a tool for the analysis of circadian phase in human sleep and chronobiological disorders,

    S. R. Pandi-Perumal, M. Smits, W. Spence, V . Srinivasan, D. P. Cardinali, A. D. Lowe, and L. Kayumov, “Dim light melatonin onset (dlmo): a tool for the analysis of circadian phase in human sleep and chronobiological disorders,”Progress in Neuro-Psychopharmacology and Biological Psychiatry, vol. 31, no. 1, pp. 1–11, 2007

    work page 2007

  9. [9]

    Comparisons of the variability of three markers of the human circadian pacemaker,

    E. B. Klerman, H. B. Gershengorn, J. F. Duffy, and R. E. Kronauer, “Comparisons of the variability of three markers of the human circadian pacemaker,”Journal of Biological Rhythms, vol. 17, no. 2, pp. 181–193, 2002

    work page 2002

  10. [10]

    Assessment of circadian rhythms,

    K. J. Reid, “Assessment of circadian rhythms,”Neurologic Clinics, vol. 37, no. 3, pp. 505–526, 2019

    work page 2019

  11. [11]

    The ingestible telemetric body core tem- perature sensor: a review of validity and exercise applications,

    C. Byrne and C. L. Lim, “The ingestible telemetric body core tem- perature sensor: a review of validity and exercise applications,”British Journal of Sports Medicine, vol. 41, no. 3, pp. 126–133, 2007

    work page 2007

  12. [12]

    Inside out: Exploring edible biocatalytic biosensors for health monitoring,

    V . Marchian `o, A. Tricase, A. Cimino, B. Cassano, M. Catacchio, E. Macchia, L. Torsi, and P. Bollella, “Inside out: Exploring edible biocatalytic biosensors for health monitoring,”Bioelectrochemistry, vol. 161, p. 108830, 2025

    work page 2025

  13. [13]

    Integrating wearable data into circadian models,

    K. M. Hannay and J. P. Moreno, “Integrating wearable data into circadian models,”Curr Opin Syst Biol, vol. 22, pp. 32–38, 2020

    work page 2020

  14. [14]

    Parameter estimation in a model of the human circadian pacemaker using a particle filter,

    J. Bonarius, C. Papatsimpa, and J. P. Linnartz, “Parameter estimation in a model of the human circadian pacemaker using a particle filter,”IEEE Transactions on Biomedical Engineering, vol. 68, no. 4, pp. 1305–1316, 2021

    work page 2021

  15. [15]

    Circadian phase prediction from non-intrusive and ambulatory physiological data,

    A. Suarez, F. Nunez, and M. Rodriguez-Fernandez, “Circadian phase prediction from non-intrusive and ambulatory physiological data,”IEEE Journal of Biomedical and Health Informatics, vol. 25, no. 5, pp. 1561– 1571, may 2021

    work page 2021

  16. [16]

    Generalizability of a neural network model for circadian phase prediction in real-world conditions,

    J. E. Stone, A. J. K. Phillips, S. Ftouni, M. Magee, M. Howard, S. W. Lockley, and S. M. W. Rajaratnam, “Generalizability of a neural network model for circadian phase prediction in real-world conditions,”Scientific Reports, vol. 9, no. 1, p. 11001, 2019

    work page 2019

  17. [17]

    Estimation of circadian body temperature rhythm based on heart rate in healthy, ambulatory subjects,

    S. Y . Sim, K. M. Joo, H. B. Kim, S. Jang, B. Kim, S. Hong, S. Kim, and K. S. Park, “Estimation of circadian body temperature rhythm based on heart rate in healthy, ambulatory subjects,”IEEE J Biomed Health Inform, vol. 21, no. 2, pp. 407–415, 2017

    work page 2017

  18. [18]

    Machine learning estimation of human body time using metabolomic profiling,

    T. Woelders, V . L. Revell, B. Middleton, K. Ackermann, M. Kayser, F. I. Raynaud, D. J. Skene, and R. A. Hut, “Machine learning estimation of human body time using metabolomic profiling,”Proceedings of the National Academy of Sciences of the United States of America, vol. 120, no. 18, 2023

    work page 2023

  19. [19]

    Bio-clock-aware of- fice lighting control,

    C. Papatsimpa, J. H. Bonarius, and J.-P. Linnartz, “Bio-clock-aware of- fice lighting control,”Proceedings of the 16th International Conference on Intelligent Environments (IE), pp. 108–114, 2020

    work page 2020

  20. [20]

    An architectural solution to a biological problem: A systematic review of lighting designs in healthcare environments,

    S. N. Hosseini, J. C. Walton, I. SheikhAnsari, N. Kreidler, and R. J. Nelson, “An architectural solution to a biological problem: A systematic review of lighting designs in healthcare environments,”Applied Sciences, vol. 14, no. 7, p. 2945, 2024

    work page 2024

  21. [21]

    Circadian rhythms: mechanisms and thera- peutic implications,

    F. Levi and U. Schibler, “Circadian rhythms: mechanisms and thera- peutic implications,”Annual Review of Pharmacology and Toxicology, vol. 47, pp. 593–628, 2007

    work page 2007

  22. [22]

    The cumulative cost of additional wakefulness: dose-response effects on neurobehavioral functions and sleep physiology from chronic sleep restriction and total sleep deprivation,

    H. P. A. Van Dongen, G. Maislin, J. M. Mullington, and D. F. Dinges, “The cumulative cost of additional wakefulness: dose-response effects on neurobehavioral functions and sleep physiology from chronic sleep restriction and total sleep deprivation,”Sleep, vol. 26, no. 2, pp. 117– 126, 2003

    work page 2003

  23. [23]

    Personalized office lighting for circadian health and improved sleep,

    C. Papatsimpa and J.-P. Linnartz, “Personalized office lighting for circadian health and improved sleep,”Sensors, vol. 20, no. 16, p. 4569, 2020

    work page 2020

  24. [24]

    Circadian rhythms in exercise performance: implications for hormonal and muscular adapta- tion,

    W. Teo, M. J. Newton, and M. R. McGuigan, “Circadian rhythms in exercise performance: implications for hormonal and muscular adapta- tion,”J Sports Sci Med, vol. 10, no. 4, pp. 600–606, 2011

    work page 2011

  25. [25]

    Chronotype, physical activity, and sport performance: a systematic review,

    J. A. Vitale and A. Weydahl, “Chronotype, physical activity, and sport performance: a systematic review,”Sports Med, vol. 47, no. 9, pp. 1859– 1868, 2017

    work page 2017

  26. [26]

    A classification approach to estimating human circadian phase under circadian alignment from actigraphy and photometry data,

    L. S. Brown, M. A. St Hilaire, A. W. McHill, A. J. K. Phillips, L. K. Barger, A. Sano, C. A. Czeisler, F. J. Doyle, and E. B. Klerman, “A classification approach to estimating human circadian phase under circadian alignment from actigraphy and photometry data,”J Pineal Res, vol. 71, no. 1, p. e12745, 2021

    work page 2021

  27. [27]

    Circadian phase estimation from ambulatory wearables with particle filtering: Accuracy depends on initialization, recording duration, and light exposure,

    L. Weed, A. Jamgochian, M. A. St Hilaire, P. Cheng, M. J. Kochender- fer, and J. M. Zeitzer, “Circadian phase estimation from ambulatory wearables with particle filtering: Accuracy depends on initialization, recording duration, and light exposure,”Journal of Biological Rhythms, 2025

    work page 2025

  28. [28]

    Computational approaches for individual circadian phase prediction in field settings,

    J. E. Stone, S. Postnova, T. L. Sletten, S. M. W. Rajaratnam, and A. J. K. Phillips, “Computational approaches for individual circadian phase prediction in field settings,”Curr Opin Syst Biol, vol. 22, pp. 39–51, 2020

    work page 2020

  29. [29]

    Predicting circadian phase across populations: a comparison of mathematical models and wearable devices,

    Y . Huang, C. Mayer, P. Cheng, A. Siddula, H. J. Burgess, C. Drake, C. Goldstein, O. Walch, and D. B. Forger, “Predicting circadian phase across populations: a comparison of mathematical models and wearable devices,”Sleep, vol. 44, no. 10, 2021

    work page 2021