pith. sign in

arxiv: 2603.15880 · v1 · submitted 2026-03-16 · 💻 cs.LG · cs.AI

Electrodermal Activity as a Unimodal Signal for Aerobic Exercise Detection in Wearable Sensors

Rena Mira Krishna , Ramya Sankar , Shadi Ghiasi This is my paper

Pith reviewed 2026-05-15 09:48 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords electrodermal activityEDAaerobic exercise detectionwearable sensorssubject-independentphasic dynamicsmachine learningactivity inference
0
0 comments X

The pith

Features from electrodermal activity alone can moderately distinguish sustained aerobic exercise from rest across different people.

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

This paper tests whether electrodermal activity, a signal reflecting sympathetic nervous system activation that is available in many wearables, can by itself separate periods of rest from sustained aerobic exercise. It applies benchmark machine learning models to EDA features from a public dataset of thirty healthy subjects and uses leave-one-subject-out validation to check generalization. The classifiers reach moderate subject-independent accuracy, with the timing and dynamics of the phasic EDA component driving much of the separation. A reader would care because this sets a clear baseline for how much information lives in this single signal before adding heart rate or motion sensors. The study frames its results as a conservative benchmark rather than a claim that EDA can stand alone in every setting.

Core claim

Using a publicly available dataset collected from thirty healthy individuals, EDA features were evaluated using benchmark machine learning models with leave-one-subject-out validation. Across models, EDA-only classifiers achieved moderate subject-independent performance, with phasic temporal dynamics and event timing contributing to class separation. Rather than proposing EDA as a replacement for multimodal sensing, this work provides a conservative benchmark of the discriminative power of EDA alone and clarifies its role as a unimodal input for wearable activity-state inference.

What carries the argument

Phasic temporal dynamics and event timing extracted from EDA signals, supplied as features to machine learning classifiers under leave-one-subject-out cross-validation.

If this is right

  • EDA can function as a unimodal input for basic activity-state inference in wearable devices.
  • Phasic temporal dynamics and event timing supply the main discriminative information between rest and exercise.
  • Subject-independent performance stays at moderate levels, indicating limits when generalizing across individuals.
  • This benchmark helps set realistic expectations for EDA before combining it with other sensors.

Where Pith is reading between the lines

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

  • Adding complementary signals such as accelerometry might raise accuracy further, though the paper does not test combinations.
  • Real-time algorithms could emphasize phasic event timing to lower power use by avoiding continuous multi-sensor sampling.
  • Testing the same approach on clinical populations or varied exercise types would reveal whether the moderate results extend beyond healthy subjects.
  • The moderate performance suggests that larger or more heterogeneous datasets could improve generalization in future work.

Load-bearing premise

The publicly available dataset from thirty healthy individuals and the derived EDA features are representative enough to support reliable subject-independent differentiation of rest from sustained aerobic exercise.

What would settle it

A new dataset from a different or larger group of subjects, run through the same EDA feature extraction and LOSO-validated models, that produces only chance-level accuracy would show the moderate performance does not hold.

Figures

Figures reproduced from arXiv: 2603.15880 by Ramya Sankar, Rena Mira Krishna, Shadi Ghiasi.

Figure 4
Figure 4. Figure 4: Confusion Matrix for [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Confusion Matrix for Linear Discriminant Analysis (LDA) [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
read the original abstract

Electrodermal Activity (EDA) is a non-invasive physiological signal widely available in wearable devices and reflects sympathetic nervous system (SNS) activation. Prior multi-modal studies have demonstrated robust performance in distinguishing stress and exercise states when EDA is combined with complementary signals such as heart rate and accelerometry. However, the ability of EDA to independently distinguish sustained aerobic exercise from low-arousal states under subject-independent evaluation remains insufficiently characterized. This study investigates whether features derived exclusively from EDA can reliably differentiate rest from sustained aerobic exercise. Using a publicly available dataset collected from thirty healthy individuals, EDA features were evaluated using benchmark machine learning models with leave-one-subject-out (LOSO) validation. Across models, EDA-only classifiers achieved moderate subject-independent performance, with phasic temporal dynamics and event timing contributing to class separation. Rather than proposing EDA as a replacement for multimodal sensing, this work provides a conservative benchmark of the discriminative power of EDA alone and clarifies its role as a unimodal input for wearable activity-state inference.

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 manuscript investigates whether features derived exclusively from electrodermal activity (EDA) can reliably differentiate rest from sustained aerobic exercise in a subject-independent manner. Using a public dataset from thirty healthy individuals, the authors extract EDA features, apply benchmark machine learning models, and evaluate them via leave-one-subject-out (LOSO) cross-validation. They report moderate subject-independent performance, with phasic temporal dynamics and event timing contributing to class separation, while framing the work as a conservative benchmark for unimodal EDA rather than a replacement for multimodal sensing.

Significance. If substantiated with quantitative metrics, this work would supply a useful reference baseline for the standalone discriminative power of EDA in wearable activity-state inference. It would help delineate the contribution of sympathetic nervous system signals to exercise detection and inform sensor-fusion designs by clarifying what unimodal EDA can and cannot achieve on its own. The use of public data and LOSO protocol supports potential replication.

major comments (2)
  1. [Abstract] Abstract: the central claim of 'moderate subject-independent performance' is stated without any numerical results (accuracy, F1, AUC, or similar), feature definitions, or error analysis. This absence prevents verification of the claim and assessment of whether phasic timing genuinely drives separation beyond what would be expected from inter-subject variability.
  2. [Evaluation] Evaluation section: LOSO on N=30 is used to support subject-independent conclusions, yet the manuscript provides no confidence intervals on the reported metrics and no external validation cohort. Given documented high between-subject differences in EDA baseline, phasic amplitude, and recovery time constants, the observed contribution of temporal dynamics may not generalize; this directly affects the load-bearing claim of reliable unimodal separation.
minor comments (2)
  1. [Methods] Methods: supply the precise mathematical definitions or references for all EDA features, especially those capturing phasic temporal dynamics and event timing, so that the feature set can be reproduced.
  2. [Discussion] Discussion: include a quantitative comparison of the achieved performance against chance level and against previously published multimodal EDA+HR+accelerometer results to contextualize the 'moderate' label.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment point by point below, indicating revisions made to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of 'moderate subject-independent performance' is stated without any numerical results (accuracy, F1, AUC, or similar), feature definitions, or error analysis. This absence prevents verification of the claim and assessment of whether phasic timing genuinely drives separation beyond what would be expected from inter-subject variability.

    Authors: We agree that the abstract would benefit from quantitative support. In the revised version, we have added the key LOSO-averaged metrics (accuracy, F1-score, and AUC with standard deviations across the 30 subjects) to substantiate the 'moderate' claim. Feature definitions are now briefly summarized in the abstract with a pointer to the Methods section, and we include a short statement on error patterns. To directly address the concern about phasic timing versus inter-subject variability, we added results from a feature-ablation analysis showing that temporal phasic features improve separation beyond tonic baseline statistics alone. revision: yes

  2. Referee: [Evaluation] Evaluation section: LOSO on N=30 is used to support subject-independent conclusions, yet the manuscript provides no confidence intervals on the reported metrics and no external validation cohort. Given documented high between-subject differences in EDA baseline, phasic amplitude, and recovery time constants, the observed contribution of temporal dynamics may not generalize; this directly affects the load-bearing claim of reliable unimodal separation.

    Authors: We have now computed and reported 95% confidence intervals for all LOSO metrics using subject-level bootstrapping to reflect the cross-validation structure. We acknowledge that an external validation cohort is not available, as the study is based exclusively on the single public dataset; this limitation is now explicitly discussed, along with the known inter-subject EDA variability. The contribution of temporal dynamics is supported by additional permutation importance and ablation results that isolate phasic event timing from baseline and amplitude effects, which we have expanded in the revised Evaluation and Discussion sections. revision: partial

standing simulated objections not resolved
  • Absence of an external validation cohort (study limited to one public dataset)

Circularity Check

0 steps flagged

No circularity: standard ML evaluation on external public dataset

full rationale

The paper applies benchmark classifiers to EDA features extracted from a publicly available dataset of 30 subjects and reports LOSO performance. No self-definitional steps exist, no fitted parameters are relabeled as predictions, and no load-bearing claims reduce to self-citations or ansatzes. The derivation chain consists of feature extraction followed by off-the-shelf model training and cross-validation on held-out subjects, remaining fully independent of the target result.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claim rests on the domain assumption that EDA reliably indexes sympathetic activation during exercise and that the 30-subject public dataset is representative for subject-independent evaluation.

axioms (1)
  • domain assumption EDA reflects sympathetic nervous system activation during exercise
    Invoked in the abstract as the physiological basis for using EDA as a unimodal signal.

pith-pipeline@v0.9.0 · 5478 in / 1057 out tokens · 36125 ms · 2026-05-15T09:48:18.283960+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

40 extracted references · 40 canonical work pages

  1. [1]

    Rena Krishna : Conceptualization, data analysis, ML modeling, manuscript drafting

    work page

  2. [2]

    Ramya Sankar : Technology review, methodological guidance, manuscript review

    work page

  3. [3]

    REFERENCES

    Shadi Ghiasi : Scientific advising, domain experience, manuscript review VIII. REFERENCES

    work page

  4. [4]

    Exercise Physiology: Nutrition, Energy, and Human Performance

    J. H. McArdle, F. I. Katch, V. L. Katch, “Exercise Physiology: Nutrition, Energy, and Human Performance ”, 9th ed. Philadelphia, PA, USA: Wolters Kluwer, 2019

    work page 2019

  5. [5]

    Exercise, GLUT4, and skeletal muscle glucose uptake,

    E. A. Richter and M. Hargreaves, “Exercise, GLUT4, and skeletal muscle glucose uptake,” Physiological Reviews , vol. 93, no. 3, pp. 993–1017, Jul. 2013

    work page 2013

  6. [6]

    A. C. Guyton and J. E. Hall, Textbook of Medical Physiology, 13th ed. Philadelphia, PA, USA: Elsevier, 2016

    work page 2016

  7. [7]

    Artificial Pancreas Control Algorithms: Challenges and Opportunities

    F. Doyle, B. Kovatchev, “Artificial Pancreas Control Algorithms: Challenges and Opportunities”, Diabetes Care, 2019

    work page 2019

  8. [8]

    Exercise and Stress Impact on Glucose Control in T1DM

    B. Buckingham, “Exercise and Stress Impact on Glucose Control in T1DM”, J. Diabetes Sci Technol, 2020

    work page 2020

  9. [9]

    The discharge behaviour of single sympathetic neurones supplying human sweat glands

    S. Ghiasi, A. Greco, R. Barbieri, E. Scilingo, G. Valenza, V. G. Macefield, and B. G. Wallin, “The discharge behaviour of single sympathetic neurones supplying human sweat glands”, Journal of the autonomic nervous system, vol. 61, no. 3, 1996

    work page 1996

  10. [10]

    Assessing Autonomic Function from Electrodermal Activity and Heart Rate Variability During Cold -Pressor Test and Emotional Challenge

    S. Ghiasi, A. Greco, R. Barbieri, E. Scilingo, G. Valenza, “ Assessing Autonomic Function from Electrodermal Activity and Heart Rate Variability During Cold -Pressor Test and Emotional Challenge ”, Scientific Reports, vol. 10, no. 1, 2020

    work page 2020

  11. [11]

    Validation of the Apple Watch for heart rate variability measurements during relax and mental stress sensors

    D. Hernando, S. Roca, J. Sancho, A. Alesanco, R. Bailón, “Validation of the Apple Watch for heart rate variability measurements during relax and mental stress sensors”, 2018

    work page 2018

  12. [12]

    L. Zhu, P. C. Ng, Y. Yu, Y. Wang, P. Spachos, D. Hatzinakos, K. N. Plataniotis, Feasibility study of Stress Detection with Machine Learning through EDA from Wearable Devices, IEEE, 2022

    work page 2022

  13. [13]

    Ishikawa, Y., Electrodermal activity and skin temperature characteristics related to stress and depression: A 4-week observational study of office workers, 2025

    work page 2025

  14. [14]

    Wearable Physiological Signals under acute Stress and Exercise Conditions

    A. Hongn, F. Bosch, L. Prado, M. P. Bonomini, “Wearable Physiological Signals under acute Stress and Exercise Conditions”, Scientific Data, vol. 12, no. 1, 2025

    work page 2025

  15. [15]

    Innovations in electrodermal activity data collection and signal processing: A systematic review,

    H. F. Posada-Quintero and K. H. Chon, “Innovations in electrodermal activity data collection and signal processing: A systematic review,” Sensors, vol. 20, no. 2, p. 479, 2020

    work page 2020

  16. [16]

    A non- eeg biosignals dataset for assessment and visualization of neurological status,

    J. Birjandtalab, D. Cogan, M. B. Pouyan, and M. Nourani, “A non- eeg biosignals dataset for assessment and visualization of neurological status,” in 2016 IEEE International Workshop on Signal Processing Systems (SiPS), pp. 110–114, IEEE, 2016

    work page 2016

  17. [17]

    Soft Wireless Bioelectronics and Differential Electrodermal Activity for Home Sleep Monitoring

    Kim, H., “Soft Wireless Bioelectronics and Differential Electrodermal Activity for Home Sleep Monitoring”, 2021

    work page 2021

  18. [18]

    Wearable Device Dataset from Induced Stress and Structured Exercise Sessions

    A. Hongn, F. Bosch, L. E. Prado, J. M. Ferrandez, M. P. Bonomini, “Wearable Device Dataset from Induced Stress and Structured Exercise Sessions ”, version 1.0.1, PhysioNet RFID: SCR_007435, 2025

    work page 2025

  19. [19]

    PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals

    A. Goldberger, L. Amaral, L. Glass, J. Hausdorff, P. C. Ivanov, R. Mark, H. E. Stanley, et al, “PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals”, 2023, pp. e215–e220. RRID:SCR_007345

    work page 2023

  20. [20]

    Exercise stress testing for primary care and sports medicine

    Harris GD. Exercise stress testing for primary care and sports medicine. In: Evans CH, White RD, editors. 1st ed. New York (NY): Springer

    work page

  21. [21]

    p. 45-54. doi: 10.1007/978-0-387-76597-6

    work page doi:10.1007/978-0-387-76597-6

  22. [22]

    Putta, Classifying Physical Stress from Relaxation Stage Using Electrodermal Activity Signal, unpublished, 2025

    A. Putta, Classifying Physical Stress from Relaxation Stage Using Electrodermal Activity Signal, unpublished, 2025

    work page 2025

  23. [23]

    On the theory of filter amplifiers,

    S. Butterworth et al. , “On the theory of filter amplifiers,” Wireless Engineer, vol. 7, no. 6, pp. 536–541, 1930

    work page 1930

  24. [24]

    Chebyshev Filters

    S. W. Smith, Digital Signal Processing: A Practical Guide for Engineers and Scientists . Elsevier. Chapter 20 "Chebyshev Filters" . 2003

    work page 2003

  25. [25]

    R. W. Schafer, What is a Savitzky -Golay filter?. IEEE Signal Processing Magazine, 28(4), 111-117, 2011

    work page 2011

  26. [26]

    P. D. Welch, The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms, IEEE Transactions on Audio and Electroacoustics, 15(2), 70–73. 1967

    work page 1967

  27. [27]

    Stress Detection Using Frequency Spectrum Analysis of Wrist - Measured Electrodermal Activity

    Stržinar, Z., Sanchis, A., Ledezma, A., Sipele, O., Pregelj, B, Škrjanc, I., “Stress Detection Using Frequency Spectrum Analysis of Wrist - Measured Electrodermal Activity”, 2023

    work page 2023

  28. [28]

    Ibrahim, S., Nazir, S., Velastin, S., Feature Selection Using Correlation Analysis and Principal Component Analysis for Accurate Breast Cancer Diagnosis, PMCID: PMC8625715, 2021

    work page 2021

  29. [29]

    -w., Jeong, J

    Chen, X. -w., Jeong, J. C., Enhanced recursive feature elimination, Sixth International Conference on Machine Learning and Applications, IEEE, 2007

    work page 2007

  30. [30]

    Faust, O., Deep learning for healthcare applications based on physiological signals: A review, Computer methods and programs in biomedicine, 2018

    work page 2018

  31. [31]

    & Liu, T

    Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., ... & Liu, T. Y. (2017). Lightgbm: A highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems, 30

    work page 2017

  32. [32]

    & Duchesnay, E., Scikit -learn: Machine learning in Python

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., ... & Duchesnay, E., Scikit -learn: Machine learning in Python. Journal of Machine Learning Research , 12(Oct), 2825-2830, 2011

    work page 2011

  33. [33]

    Bent, B., Investigating sources of inaccuracy in wearable optical heart rate sensors, NPJ Digital Medicine, 3, article 18, 2020

    work page 2020

  34. [34]

    Lundberg, S., Lee, S., A Unified Appraoch to Interpreting Model Predictions, Neurips, 2017

    work page 2017

  35. [35]

    (2016, August)

    Chen, T., & Guestrin, C. (2016, August). Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining (pp. 785-794)

    work page 2016

  36. [36]

    Raschka, S., Mirjalili, V., Python Machine Learning: Machine Learning and Deep Learning with Python, scikit -learn, and TensorFlow 2 (3rd ed.), 2019

    work page 2019

  37. [37]

    Heart rate variability, qt variability, and electrodermal activity during exercise,

    S. Boettger, C. Puta, V. K. Yeragani, L. Donath, H. -J. Mueller, H. H. Gabriel, and K. -J. Baer, “Heart rate variability, qt variability, and electrodermal activity during exercise,” Med Sci Sports Exerc, vol. 42, no. 3, pp. 443–8, 2010

    work page 2010

  38. [38]

    Buckingham, K

    S. Buckingham, K. Morrisey, A. Williams, L. Price, J. harrison, The Physical Activity Wearables in the Police Force (PAW -Force) study: acceptability and impact, Springer, 2020

    work page 2020

  39. [39]

    P. Cho, J. Kim, B. Bent, J. Dunn, BIG IDEAs Lab Glycemic Variability and Wearable Device Data, Physionet, version 1.1.2, 2023

    work page 2023

  40. [40]

    N. Khan, N. Deshpande, P. Kadam, A review of AI -based health prediction using apple watch and fitbit data, Springer, 2025

    work page 2025