arxiv: 2605.05246 · v1 · submitted 2026-05-04 · 📡 eess.SP · cs.AI

Recognition: unknown

Memory-Efficient EDA Denoising via Knowledge Distillation for Wearable IoT Under Severe Motion Artifacts and Underwater Conditions

Andrew Peitzsch, Dong-hee Kang, Farnoush Baghestani, Jarod Zizza, Ki H. Chon, Yongbin Lee, Youngsun Kong

Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:32 UTC · model grok-4.3

classification 📡 eess.SP cs.AI

keywords electrodermal activityknowledge distillationdenoisingwearable devicesmotion artifactsunderwaterskin conductance responsehealth monitoring

0 comments

The pith

A knowledge-distilled lightweight CNN model denoises EDA signals effectively under motion artifacts and underwater conditions while reducing size and compute by over 90%.

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

This paper aims to create a denoising system for electrodermal activity that can run on wearable devices even when users are underwater or moving intensely. It trains a small student network by distilling knowledge from a larger hybrid CNN and Transformer teacher model, using specially augmented data to represent real noise. The result is a model that cleans the signals well enough to support accurate health monitoring and early warnings for conditions like oxygen toxicity. Such a system would make reliable continuous monitoring possible in field settings where current methods fail due to noise or hardware limits.

Core claim

The paper claims that integrating a hybrid CNN-Transformer teacher model with a lightweight depth-wise separable CNN student model through knowledge distillation, combined with a realistic data augmentation scheme for motion artifacts and environmental distortions, produces a deployable denoiser that preserves performance metrics like MAE of 0.144 and SNR improvement of 12.08 dB, substantially improves reconstruction on real underwater data, and enhances downstream clinical predictions.

What carries the argument

Knowledge distillation transferring capabilities from the teacher to the student model, with the student being a depth-wise separable CNN and training aided by realistic simulation of artifacts and distortions.

If this is right

Model storage drops from 7.87 MB to 0.51 MB and computations from 105.1M to 11.61M FLOPs, fitting resource-constrained wearables.
Underwater skin conductance reconstruction error falls from 2.809 to 0.215 on the UMAC dataset.
CNS-OT prediction AUROC reaches 0.806 with sensitivity improving to 0.767, enabling predictions a median of 6.9 minutes earlier.
The approach generalizes across measurement locations and harsh environments in validation tests.

Where Pith is reading between the lines

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

This compression method might extend to real-time denoising of other biosignals in mobile health applications where power and memory are limited.
The early prediction improvement suggests potential for proactive interventions in high-risk activities if further clinical validation confirms the gains.
Deployment on actual IoT hardware could reveal additional optimizations or trade-offs not captured in the simulated tests.

Load-bearing premise

The data augmentation accurately captures the real statistical distribution of motion artifacts and underwater distortions so performance transfers to actual harsh conditions without significant domain shift.

What would settle it

Testing the student model on a new collection of real underwater EDA signals from different subjects and devices, where it fails to reduce reconstruction error or improve prediction metrics beyond undenoised signals, would falsify the generalization.

Figures

Figures reproduced from arXiv: 2605.05246 by Andrew Peitzsch, Dong-hee Kang, Farnoush Baghestani, Jarod Zizza, Ki H. Chon, Yongbin Lee, Youngsun Kong.

**Figure 1.** Figure 1: Comparison of EDA signals across different measurement locations view at source ↗

**Figure 3.** Figure 3: Representative examples of the noise-augmented signal, clean target, and real MA-corrupted EDA collected simultaneously. (a) Timedomain comparison, demonstrating that the noise-augmented signal mimics real MA-corrupted EDA while preserving the underlying baseline dynamics of the clean EDA, in contrast to the asymmetric real MAcorrupted signal. (b) Power spectral density comparison showing that the augmen… view at source ↗

**Figure 4.** Figure 4: Overview of the proposed EDA denoising framework. (a) Teacher model based on a hybrid CNN view at source ↗

**Figure 5.** Figure 5: An example comparison of EDA denoising results on a signal view at source ↗

read the original abstract

Electrodermal activity (EDA) is widely used in wearable Internet of Medical Things (IoMT) systems for continuous health monitoring, including autonomic assessment. However, EDA signals are highly vulnerable to motion artifacts and environmental noise, limiting reliable deployment in harsh operating conditions such as underwater. This study proposes a robust, deployable EDA denoising framework that generalizes across multiple measurement locations and harsh environments. The framework integrates a hybrid CNN-Transformer teacher model with a lightweight depth-wise separable CNN student model via a knowledge distillation (KD) strategy. To further improve robustness, a realistic data augmentation scheme is introduced to simulate diverse motion artifacts and environmental distortions. The KD-based student model significantly reduces model size (7.87 MB to 0.51 MB) and computational cost (105.1M to 11.61M FLOPs) while maintaining denoising performance (MAE: 0.144, SNR improvement: 12.08 dB) using the public dataset validation. In real-world underwater conditions (UMAC dataset) testing, the proposed method substantially improves skin conductance response reconstruction, reducing mean absolute error from 2.809 to 0.215. Furthermore, on independent testing using the CNS-OT dataset, the denoised signals enhanced downstream CNS-OT prediction performance, achieving the highest AUROC (0.806) compared to prior denoising methods. The proposed method also improved the early prediction rate (sensitivity) from 0.550 to 0.767, enabling CNS-OT prediction up to a median of 6.9 minutes before symptom onset. These results demonstrate that the proposed framework not only improves EDA signal quality but also enhances clinically relevant prediction performance while remaining suitable for deployment in resource-constrained wearable Internet of Things systems operating in harsh environments.

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.

Desk Editor's Note private letter to a colleague

The distillation shrinks the EDA denoiser enough for wearables and shows clear gains on real underwater recordings plus better CNS-OT prediction, but the augmentation's match to actual conditions is not checked.

read the letter

Colleague, the core result is that a hybrid CNN-Transformer teacher can be distilled into a depth-wise separable CNN student that cuts model size from 7.87 MB to 0.51 MB and FLOPs from 105M to 11.6M while holding denoising performance at MAE 0.144 and 12 dB SNR gain on public validation data. On the independent UMAC underwater set the student drops MAE from 2.809 to 0.215, and on CNS-OT it lifts AUROC to 0.806 and sensitivity to 0.767, moving the median early-prediction window to 6.9 minutes. That pipeline—teacher architecture, KD transfer, and domain-specific augmentation—is what is actually new here, and the direct tie from cleaned signals to a downstream clinical task is the part that matters for deployment. The numbers are reported on held-out and real-world sets rather than circular fits, which is a plus. The soft spot is the augmentation step. The method trains on simulated motion artifacts and underwater distortions then evaluates on genuine UMAC recordings, yet the paper gives no quantitative evidence (spectral overlap, distribution distances, or cross-validation against held-out real noise) that the simulations reproduce the target statistics. Without that check the large reported lift on UMAC could partly reflect domain mismatch rather than true robustness. Baseline comparisons and statistical tests are also thin in the write-up. This work is for groups building resource-constrained wearables for sports, military, or marine monitoring who need concrete efficiency numbers and real-environment results. It is coherent on its own terms and has enough reproducible claims to deserve a serious referee, even though reviewers will need to press on the augmentation validation and fuller ablation details. I would send it to peer review.

Referee Report

2 major / 1 minor

Summary. The paper claims to develop a memory-efficient EDA denoising method using knowledge distillation from a hybrid CNN-Transformer teacher to a lightweight depth-wise separable CNN student, augmented with a realistic data augmentation scheme for motion artifacts and underwater distortions. On public validation, it achieves MAE of 0.144 and SNR improvement of 12.08 dB with reduced model size (0.51 MB) and FLOPs (11.61M). On the UMAC underwater dataset, it reduces MAE from 2.809 to 0.215 for skin conductance response reconstruction. On the CNS-OT dataset, it achieves AUROC of 0.806 and improves sensitivity from 0.550 to 0.767 for early prediction up to 6.9 minutes before onset.

Significance. If the results hold, this framework could significantly advance wearable IoMT systems by enabling reliable EDA-based monitoring in challenging environments such as underwater, where motion artifacts and environmental noise are severe. The model compression aspect is particularly valuable for resource-limited devices, and the demonstrated improvements in downstream clinical prediction tasks (CNS-OT) add practical significance beyond signal quality metrics. The approach addresses a real gap in deploying EDA in harsh conditions.

major comments (2)

[Data Augmentation and Generalization] The claimed generalization to real-world UMAC underwater conditions (MAE reduction from 2.809 to 0.215) depends on the fidelity of the proposed data augmentation scheme in simulating motion artifacts and environmental distortions. However, the manuscript provides no quantitative validation, such as statistical distribution comparisons, spectral analysis, or domain adaptation metrics, to confirm that the augmented training data matches the target domain's joint statistics. This is a load-bearing assumption for the transfer performance claims.
[Experimental Evaluation] The reported performance metrics (e.g., AUROC 0.806, sensitivity improvement) lack accompanying details on the specific baseline denoising methods used for comparison, statistical tests for significance of improvements, ablation studies isolating the contributions of KD and augmentation, and full experimental protocols including cross-validation procedures and hyperparameter tuning. These omissions hinder assessment of the reliability and reproducibility of the central claims.

minor comments (1)

[Abstract] The abstract mentions 'public dataset validation' but does not specify which public datasets were used for training and validation, which would aid clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help strengthen the manuscript. We address each major comment below and will incorporate revisions to improve rigor and reproducibility.

read point-by-point responses

Referee: [Data Augmentation and Generalization] The claimed generalization to real-world UMAC underwater conditions (MAE reduction from 2.809 to 0.215) depends on the fidelity of the proposed data augmentation scheme in simulating motion artifacts and environmental distortions. However, the manuscript provides no quantitative validation, such as statistical distribution comparisons, spectral analysis, or domain adaptation metrics, to confirm that the augmented training data matches the target domain's joint statistics. This is a load-bearing assumption for the transfer performance claims.

Authors: We agree that explicit quantitative validation of the augmentation fidelity is necessary to support the generalization claims to the UMAC dataset. In the revised manuscript, we will add a dedicated analysis subsection including: Kolmogorov-Smirnov tests on signal statistics (mean, variance, peak amplitude); power spectral density comparisons between augmented and real underwater signals; and domain adaptation metrics such as Maximum Mean Discrepancy (MMD) and Fréchet Inception Distance on extracted features. These results will be reported with p-values to substantiate that the augmented data sufficiently approximates the target domain's joint statistics. revision: yes
Referee: [Experimental Evaluation] The reported performance metrics (e.g., AUROC 0.806, sensitivity improvement) lack accompanying details on the specific baseline denoising methods used for comparison, statistical tests for significance of improvements, ablation studies isolating the contributions of KD and augmentation, and full experimental protocols including cross-validation procedures and hyperparameter tuning. These omissions hinder assessment of the reliability and reproducibility of the central claims.

Authors: We acknowledge these omissions limit reproducibility assessment. The revised manuscript will expand the Experimental Setup and Results sections to: (1) explicitly list and reference all baseline denoising methods with implementation details; (2) include statistical significance tests (paired t-tests for MAE/SNR with p-values, McNemar's test for AUROC/sensitivity); (3) present full ablation studies (e.g., teacher-only, student without KD, with/without augmentation); and (4) detail protocols including 5-fold cross-validation, hyperparameter search ranges, early stopping criteria, and training hyperparameters. These additions will be placed in the main text and supplementary material. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical results on independent datasets

full rationale

The paper's core claims rest on empirical measurements of denoising performance (MAE, SNR) and downstream task improvements (AUROC, sensitivity) obtained on held-out public datasets plus independent real-world UMAC and CNS-OT recordings. The knowledge-distillation architecture and data-augmentation scheme are methodological proposals whose outputs are evaluated externally; no equations, fitted parameters, or self-citations reduce the reported gains to definitions or tautologies of the inputs. The augmentation fidelity assumption affects generalization validity but does not create a circular derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework rests on standard supervised learning assumptions plus the unproven transferability of simulated noise to real harsh environments; no new physical entities or ad-hoc constants are introduced.

axioms (2)

domain assumption Knowledge distillation transfers denoising capability from teacher to student without critical loss of signal features needed for downstream tasks.
Invoked by the KD strategy that trains the student to match teacher outputs on augmented data.
domain assumption Synthetic motion and environmental distortions generated by the augmentation scheme have the same statistical effect on EDA as real-world artifacts.
Central to the claim that training on augmented data generalizes to UMAC underwater recordings.

pith-pipeline@v0.9.0 · 5658 in / 1519 out tokens · 120476 ms · 2026-05-08T17:32:50.800722+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

39 extracted references · 37 canonical work pages · 3 internal anchors

[1]

Boucsein, Electrodermal Activity

W. Boucsein, Electrodermal Activity. Springer Science & Business Media, 2012

2012
[2]

An Integrated Wearable Sensor for Unobtrusive Continuous Measurement of Autonomic Nervous System,

M. S. Mahmud, H. Fang, and H. Wang, “An Integrated Wearable Sensor for Unobtrusive Continuous Measurement of Autonomic Nervous System,” IEEE Internet Things J., vol. 6, no. 1, pp. 1104– 1113, Feb. 2019, doi: 10.1109/JIOT.2018.2868235

work page doi:10.1109/jiot.2018.2868235 2019
[3]

MDNet: A Lightweight Multi-Domain 1D-CNN for Embedded Pain Assessment Using EDA Signals,

S. Aziz, G. Chetty, R. Goecke, and R. F. Rojas, “MDNet: A Lightweight Multi-Domain 1D-CNN for Embedded Pain Assessment Using EDA Signals,” IEEE Internet Things J., pp. 1–1, 2026, doi: 10.1109/JIOT.2026.3663683

work page doi:10.1109/jiot.2026.3663683 2026
[4]

Toward Stress- Adaptive Cyber Defense: Cognitive–Physiological Synchronization in IoT Environments,

A. Yazdinejad, H. Karimipour, and T. Halabi, “Toward Stress- Adaptive Cyber Defense: Cognitive–Physiological Synchronization in IoT Environments,” IEEE Internet Things J., vol. 13, no. 7, pp. 13832–13848, Apr. 2026, doi: 10.1109/JIOT.2026.3656466

work page doi:10.1109/jiot.2026.3656466 2026
[5]

Electrodermal activity in pain assessment and its clinical applications,

Y. Kong and K. H. Chon, “Electrodermal activity in pain assessment and its clinical applications,” Appl. Phys. Rev., vol. 11, no. 3, p. 031316, Aug. 2024, doi: 10.1063/5.0200395

work page doi:10.1063/5.0200395 2024
[6]

Twenty-four-hour patterns in electrodermal activity recordings of patients with and without epileptic seizures,

S. Vieluf et al., “Twenty-four-hour patterns in electrodermal activity recordings of patients with and without epileptic seizures,” Epilepsia, vol. 62, no. 4, pp. 960–972, 2021, doi: 10.1111/epi.16843

work page doi:10.1111/epi.16843 2021
[7]

Seizures Caused by Exposure to Hyperbaric Oxygen in Rats Can Be Predicted by Early Changes in Electrodermal Activity,

H. F. Posada-Quintero, C. S. Landon, N. M. Stavitzski, J. B. Dean, and K. H. Chon, “Seizures Caused by Exposure to Hyperbaric Oxygen in Rats Can Be Predicted by Early Changes in Electrodermal Activity,” Front. Physiol., vol. 12, Jan. 2022, doi: 10.3389/fphys.2021.767386

work page doi:10.3389/fphys.2021.767386 2022
[8]

Oxygen Toxicity and Special Operations Forces Diving: Hidden and Dangerous,

T. T. Wingelaar, P.-J. A. M. van Ooij, and R. A. van Hulst, “Oxygen Toxicity and Special Operations Forces Diving: Hidden and Dangerous,” Front. Psychol., vol. 8, Jul. 2017, doi: 10.3389/fpsyg.2017.01263

work page doi:10.3389/fpsyg.2017.01263 2017
[9]

Elevation of spectral components of electrodermal activity precedes central nervous system oxygen toxicity symptoms in divers,

H. F. Posada–Quintero et al., “Elevation of spectral components of electrodermal activity precedes central nervous system oxygen toxicity symptoms in divers,” Commun Med, vol. 4, no. 1, p. 270, Dec. 2024, doi: 10.1038/s43856-024-00688-4

work page doi:10.1038/s43856-024-00688-4 2024
[10]

Prediction of central nervous system oxygen toxicity symptoms using electrodermal activity and machine learning,

M.-B. Hossain et al., “Prediction of central nervous system oxygen toxicity symptoms using electrodermal activity and machine learning,” Biocybernetics and Biomedical Engineering, vol. 44, no. 2, pp. 304–311, Apr. 2024, doi: 10.1016/j.bbe.2024.03.004

work page doi:10.1016/j.bbe.2024.03.004 2024
[11]

Unsupervised motion artifact detection in wrist-measured electrodermal activity data,

Y. Zhang, M. Haghdan, and K. S. Xu, “Unsupervised motion artifact detection in wrist-measured electrodermal activity data,” in Proceedings of the 2017 ACM International Symposium on Wearable Computers, in ISWC ’17. New York, NY, USA: Association for Computing Machinery, Sep. 2017, pp. 54–57. doi: 10.1145/3123021.3123054

work page doi:10.1145/3123021.3123054 2017
[12]

Automatic motion artifact detection in electrodermal activity data using machine learning,

M.-B. Hossain, H. F. Posada-Quintero, Y. Kong, R. McNaboe, and K. H. Chon, “Automatic motion artifact detection in electrodermal activity data using machine learning,” Biomed. Signal Process. Control, vol. 74, p. 103483, Apr. 2022, doi: 10.1016/j.bspc.2022.103483

work page doi:10.1016/j.bspc.2022.103483 2022
[13]

Automatic motion artifact detection in electrodermal activity signals using 1D U-net architecture,

Y. Kong, M. B. Hossain, A. Peitzsch, H. F. Posada-Quintero, and K. H. Chon, “Automatic motion artifact detection in electrodermal activity signals using 1D U-net architecture,” Comput. Biol. Med., vol. 182, p. 109139, Nov. 2024, doi: 10.1016/j.compbiomed.2024.109139

work page doi:10.1016/j.compbiomed.2024.109139 2024
[14]

Wavelet-based motion artifact removal for electrodermal activity,

W. Chen, N. Jaques, S. Taylor, A. Sano, S. Fedor, and R. W. Picard, “Wavelet-based motion artifact removal for electrodermal activity,” in 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Aug. 2015, pp. 6223–6226. doi: 10.1109/EMBC.2015.7319814

work page doi:10.1109/embc.2015.7319814 2015
[15]

Efficient wavelet-based artifact removal for electrodermal activity in real-world applications,

J. Shukla, M. Barreda-Ángeles, J. Oliver, and D. Puig, “Efficient wavelet-based artifact removal for electrodermal activity in real-world applications,” Biomed. Signal Process. Control, vol. 42, pp. 45–52, Apr. 2018, doi: 10.1016/j.bspc.2018.01.009

work page doi:10.1016/j.bspc.2018.01.009 2018
[16]

A Deep Convolutional Autoencoder for Automatic Motion Artifact Removal in Electrodermal Activity,

M.-B. Hossain, H. F. Posada-Quintero, and K. H. Chon, “A Deep Convolutional Autoencoder for Automatic Motion Artifact Removal in Electrodermal Activity,” IEEE Trans. Biomed. Eng., vol. 69, no. 12, pp. 3601–3611, Dec. 2022, doi: 10.1109/TBME.2022.3174509

work page doi:10.1109/tbme.2022.3174509 2022
[17]

Comparison of Electrodermal Activity from Multiple Body Locations Based on Standard EDA Indices’ Quality and Robustness against Motion Artifact,

M.-B. Hossain, Y. Kong, H. F. Posada-Quintero, and K. H. Chon, “Comparison of Electrodermal Activity from Multiple Body Locations Based on Standard EDA Indices’ Quality and Robustness against Motion Artifact,” Sensors, vol. 22, no. 9, p. 3177, Jan. 2022, doi: 10.3390/s22093177

work page doi:10.3390/s22093177 2022
[18]

Bodily Electrodermal Representations for Affective Computing,

X. Shui et al., “Bodily Electrodermal Representations for Affective Computing,” IEEE Trans. Affect. Comput., vol. 15, no. 3, pp. 1018– 1025, Jul. 2024, doi: 10.1109/TAFFC.2023.3315973

work page doi:10.1109/taffc.2023.3315973 2024
[19]

Sex differences in autonomic functions and cognitive performance during cold-air exposure and cold-water partial immersion,

Y. Kong et al., “Sex differences in autonomic functions and cognitive performance during cold-air exposure and cold-water partial immersion,” Front. Physiol., vol. 15, Oct. 2024, doi: 10.3389/fphys.2024.1463784

work page doi:10.3389/fphys.2024.1463784 2024
[20]

Multiple Arousal Theory and Daily-Life Electrodermal Activity Asymmetry,

R. W. Picard, S. Fedor, and Y. Ayzenberg, “Multiple Arousal Theory and Daily-Life Electrodermal Activity Asymmetry,” Emotion Review, vol. 8, no. 1, pp. 62–75, Jan. 2016, doi: 10.1177/1754073914565517

work page doi:10.1177/1754073914565517 2016
[21]

MCUNet: Tiny Deep Learning on IoT Devices,

J. Lin, W.-M. Chen, Y. Lin, J. Cohn, C. Gan, and S. Han, “MCUNet: Tiny Deep Learning on IoT Devices,” in Advances in Neural Information Processing Systems, 2020, pp. 11711–11722

2020
[22]

Distilling the Knowledge in a Neural Network

G. Hinton, O. Vinyals, and J. Dean, “Distilling the Knowledge in a Neural Network,” Mar. 09, 2015, arXiv: 1503.02531. doi: 10.48550/arXiv.1503.02531

work page Pith review doi:10.48550/arxiv.1503.02531 2015
[23]

FiLM : Visual reasoning with a general conditioning layer

E. Perez, F. Strub, H. de Vries, V. Dumoulin, and A. Courville, “FiLM: Visual Reasoning with a General Conditioning Layer,” 13 This work has been submitted to IEEE for possible publication. Copyright may be transferred without notice. Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32, no. 1, Apr. 2018, doi: 10.1609/aaai.v32i1.11671

work page doi:10.1609/aaai.v32i1.11671 2018
[24]

Modelling event-related skin conductance responses,

D. R. Bach, G. Flandin, K. J. Friston, and R. J. Dolan, “Modelling event-related skin conductance responses,” International Journal of Psychophysiology, vol. 75, no. 3, pp. 349–356, Mar. 2010, doi: 10.1016/j.ijpsycho.2010.01.005

work page doi:10.1016/j.ijpsycho.2010.01.005 2010
[25]

Time-series analysis for rapid event-related skin conductance responses,

D. R. Bach, G. Flandin, K. J. Friston, and R. J. Dolan, “Time-series analysis for rapid event-related skin conductance responses,” Journal of Neuroscience Methods, vol. 184, no. 2, pp. 224–234, Nov. 2009, doi: 10.1016/j.jneumeth.2009.08.005

work page doi:10.1016/j.jneumeth.2009.08.005 2009
[26]

Dynamic causal modelling of anticipatory skin conductance responses,

D. R. Bach, J. Daunizeau, K. J. Friston, and R. J. Dolan, “Dynamic causal modelling of anticipatory skin conductance responses,” Biological Psychology, vol. 85, no. 1, pp. 163–170, Sep. 2010, doi: 10.1016/j.biopsycho.2010.06.007

work page doi:10.1016/j.biopsycho.2010.06.007 2010
[27]

Optimising a model-based approach to inferring fear learning from skin conductance responses,

M. Staib, G. Castegnetti, and D. R. Bach, “Optimising a model-based approach to inferring fear learning from skin conductance responses,” Journal of Neuroscience Methods, vol. 255, pp. 131–138, Nov. 2015, doi: 10.1016/j.jneumeth.2015.08.009

work page doi:10.1016/j.jneumeth.2015.08.009 2015
[28]

Assessing fear learning via conditioned respiratory amplitude responses,

G. Castegnetti, A. Tzovara, M. Staib, S. Gerster, and D. R. Bach, “Assessing fear learning via conditioned respiratory amplitude responses,” Psychophysiology, vol. 54, no. 2, pp. 215–223, 2017, doi: 10.1111/psyp.12778

work page doi:10.1111/psyp.12778 2017
[29]

Modeling startle eyeblink electromyogram to assess fear learning,

S. Khemka, A. Tzovara, S. Gerster, B. B. Quednow, and D. R. Bach, “Modeling startle eyeblink electromyogram to assess fear learning,” Psychophysiology, vol. 54, no. 2, pp. 204–214, 2017, doi: 10.1111/psyp.12775

work page doi:10.1111/psyp.12775 2017
[30]

A pupil size response model to assess fear learning,

C. W. Korn, M. Staib, A. Tzovara, G. Castegnetti, and D. R. Bach, “A pupil size response model to assess fear learning,” Psychophysiology, vol. 54, no. 3, pp. 330–343, 2017, doi: 10.1111/psyp.12801

work page doi:10.1111/psyp.12801 2017
[31]

Temporally Unpredictable Sounds Exert a Context-Dependent Influence on Evaluation of Unrelated Images,

D. R. Bach, E. Seifritz, and R. J. Dolan, “Temporally Unpredictable Sounds Exert a Context-Dependent Influence on Evaluation of Unrelated Images,” PLOS ONE, vol. 10, no. 6, p. e0131065, Jun. 2015, doi: 10.1371/journal.pone.0131065

work page doi:10.1371/journal.pone.0131065 2015
[32]

A head-to-head comparison of SCRalyze and Ledalab, two model-based methods for skin conductance analysis,

D. R. Bach, “A head-to-head comparison of SCRalyze and Ledalab, two model-based methods for skin conductance analysis,” Biological Psychology, vol. 103, pp. 63–68, Dec. 2014, doi: 10.1016/j.biopsycho.2014.08.006

work page doi:10.1016/j.biopsycho.2014.08.006 2014
[33]

Mechanically Robust Superhydrophobic Coatings via Dual-Step Deposition for Electrodermal Activity (EDA) Electrodes Immersible in Saltwater,

D. Kang et al., “Mechanically Robust Superhydrophobic Coatings via Dual-Step Deposition for Electrodermal Activity (EDA) Electrodes Immersible in Saltwater,” Apr. 24, 2026, Social Science Research Network, Rochester, NY: 6641936. doi: 10.2139/ssrn.6641936

work page doi:10.2139/ssrn.6641936 2026
[34]

TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation

J. Chen et al., “TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation,” Feb. 08, 2021, arXiv: 2102.04306. doi: 10.48550/arXiv.2102.04306

work page internal anchor Pith review doi:10.48550/arxiv.2102.04306 2021
[35]

Atrial Fibrillation Prediction Using a Lightweight Temporal Convolutional and Selective State Space Architecture,

Y. Lee and K. H. Chon, “Atrial Fibrillation Prediction Using a Lightweight Temporal Convolutional and Selective State Space Architecture,” Aug. 26, 2025, arXiv: 2508.19361. doi: 10.48550/arXiv.2508.19361

work page doi:10.48550/arxiv.2508.19361 2025
[36]

MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

A. G. Howard et al., “MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,” Apr. 17, 2017, arXiv: 1704.04861. doi: 10.48550/arXiv.1704.04861

work page internal anchor Pith review doi:10.48550/arxiv.1704.04861 2017
[37]

cvxEDA: A Convex Optimization Approach to Electrodermal Activity Processing,

A. Greco, G. Valenza, A. Lanata, E. P. Scilingo, and L. Citi, “cvxEDA: A Convex Optimization Approach to Electrodermal Activity Processing,” IEEE Trans. Biomed. Eng., vol. 63, no. 4, pp. 797–804, Apr. 2016, doi: 10.1109/TBME.2015.2474131

work page doi:10.1109/tbme.2015.2474131 2016
[38]

ospEDA: Orthogonal Subspace Projection for Electrodermal Activity Decomposition

Y. Lee, Y. Kong, and K. H. Chon, “ospEDA: Orthogonal Subspace Projection for Electrodermal Activity Decomposition,” Apr. 08, 2026, arXiv: 2604.07521. doi: 10.48550/arXiv.2604.07521

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.07521 2026
[39]

Call Center Stress Recognition with Person-Specific Models,

J. Hernandez, R. R. Morris, and R. W. Picard, “Call Center Stress Recognition with Person-Specific Models,” in Affective Computing and Intelligent Interaction, S. D’Mello, A. Graesser, B. Schuller, and J.-C. Martin, Eds., Berlin, Heidelberg: Springer, 2011, pp. 125–134. doi: 10.1007/978-3-642-24600-5_16

work page doi:10.1007/978-3-642-24600-5_16 2011