arxiv: 2604.24004 · v1 · submitted 2026-04-27 · 📡 eess.SP

Recognition: unknown

Single-Cycle Multidirectional EOG Classification Faster than Human Reaction Time for Wearable Human-Computer Interactions

Tasnia Nabiha , Orthy Toor , Wakim Sajjad Sakib , Abdullah Bin Shams

Authors on Pith no claims yet

Pith reviewed 2026-05-08 02:23 UTC · model grok-4.3

classification 📡 eess.SP

keywords latencysingle-cycleclassificationcascadedhumanreactionsignalsaccuracy

0 comments

The pith

Cascaded neural networks classify 10 eye-movement classes from single-cycle EOG signals at 99% accuracy with sub-83 ms latency below human reaction time.

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

Electrooculogram signals are tiny voltage changes picked up near the eyes that reflect where someone is looking or if they blink. The work focuses on using only a short single cycle of these signals instead of longer recordings. This choice reduces delay and power use but leaves less information for the computer to work with and makes noise a bigger problem. The authors trained two kinds of neural networks: a straightforward one-dimensional network and a cascaded setup that combines simpler and more complex layers. Both were tested on ten possible actions including staring, blinking, and looking in eight diagonal and cardinal directions. The models reached roughly 99 percent correct classification while responding in 39 to 83 milliseconds, faster than the roughly 200 milliseconds most people need to react to something they see. Explainable AI techniques were used to inspect what parts of the short signal the networks relied on. The approach is aimed at wearable devices where quick, accurate eye commands could replace buttons or touchscreens.

Core claim

The study achieved an accuracy of around 99% for all the models with a latency of 38.6 ms for the 1D ANN, and 82.85 ms for the cascaded CNN.

Load-bearing premise

That the high accuracy observed on the collected single-cycle EOG data will generalize to new users, varying electrode placements, and real-world noise levels without substantial retraining or post-hoc data selection.

Figures

Figures reproduced from arXiv: 2604.24004 by Abdullah Bin Shams, Orthy Toor, Tasnia Nabiha, Wakim Sajjad Sakib.

**Figure 12.** Figure 12: Linear Discriminant Analysis (LDA) projection of the 26- dimensional feature space onto the first two discriminant components. The visualization illustrates enhanced class separability compared to PCA due to supervised variance maximization between eye-movement categories. Linear Discriminant Analysis (LDA) as shown in view at source ↗

read the original abstract

Electrooculogram (EOG) is a non-invasive bio-signal generated by the potential difference between the retina and cornea during eye movement, and is widely utilized in Human-Computer Interaction (HCI) systems. Expanding the range of detectable eye movements enhances system capability. However, increasing the number of classes typically degrades classification performance. While AI-based approaches can mitigate this limitation, their complexity increases significantly when operating on single-cycle EOG signals. Although single-cycle signals offer advantages such as low latency, reduced power consumption, and improved responsiveness, they are inherently limited by reduced informational content and higher susceptibility to noise. Ensuring low latency remains critical for real-time HCI applications, where system response must remain below human reaction thresholds. In this experimental study, using explainable AI, we address these challenges by developing 1-dimensional (1D) and cascaded ANN and CNN architectures capable of highly accurate classification across ten EOG classes (Stare, Blink, Up, Down, Right, Left, Up-left, Up-right, Down-left, and Down-right) using single-cycle signals, while simultaneously achieving latency substantially lower than human reaction time. The study achieved an accuracy of around 99% for all the models with a latency of 38.6 ms for the 1D ANN, and 82.85 ms for the cascaded CNN. These findings confirm that cascaded neural network architectures, can effectively balance high classification accuracy and low latency for single-cycle, multi-class EOG-based HCI systems under limited data availability.

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

They built a 10-class single-cycle EOG classifier that hits reported 99% accuracy and sub-100 ms latency with basic ANN/CNN models, but the validation details are missing so the numbers are hard to trust.

read the letter

The paper's main result is a concrete engineering demo: ten EOG classes (including diagonals) classified from single-cycle signals using 1D ANN and cascaded CNN setups, with latencies of 38.6 ms and 82.85 ms respectively and accuracy around 99%. They frame it as solving the accuracy-latency trade-off for wearable HCI while staying under human reaction time, and they bring in explainable AI to look at what the models use. That combination of class count, single-cycle constraint, and cascaded architecture is a reasonable incremental step from earlier EOG work on fewer classes or longer windows. The numbers are specific enough that someone could try to reproduce the architectures on their own data. The soft spot is exactly what the stress-test flagged. The abstract gives no dataset size, no subject count, no mention of cross-subject or leave-one-out splits, and no noise model or baseline comparisons. Single-cycle EOG is short and noisy by nature, so 99% accuracy is easy to get inside one person's clean recordings but usually falls apart with new users, slight electrode movement, or everyday interference. Without those checks the central claim does not yet hold up for real deployment. This is the sort of applied paper that might interest people building hands-free interfaces for wearables or assistive devices who need low-latency multi-class eye control. A reader looking for ready-to-try network shapes and latency targets could pull useful pieces from it. I would send it to peer review because the practical goal is clear and the reported latencies are worth checking, but the authors would need to add proper inter-subject validation and data details before it could be taken as reliable evidence.

Referee Report

2 major / 1 minor

Summary. The paper develops 1D ANN and cascaded CNN architectures for classifying ten single-cycle EOG classes (stare, blink, up, down, left, right, up-left, up-right, down-left, down-right). It reports ~99% accuracy for all models with latencies of 38.6 ms (1D ANN) and 82.85 ms (cascaded CNN), claiming these are below human reaction time and suitable for low-power wearable HCI under limited data, aided by explainable AI.

Significance. If the performance claims hold under proper validation, the work would demonstrate that cascaded neural architectures can deliver multi-class, sub-100 ms EOG classification from single-cycle signals, offering practical advantages in latency, power, and responsiveness for wearable interfaces. The explicit focus on single-cycle operation and latency benchmarking against human reaction time is a useful framing for real-time HCI.

major comments (2)

[Abstract] Abstract: the central claim of ~99% accuracy on 10 classes with sub-100 ms latency is presented without any dataset size, number of subjects, train/test split, cross-validation procedure (intra- vs. inter-subject), noise model, or error bars. Single-cycle EOG is known to be noise-sensitive; absent these details the reported numbers cannot be assessed for robustness or statistical significance.
[Abstract] Abstract: no leave-one-subject-out, inter-subject, or cross-session metrics are mentioned. The performance therefore rests on the untested assumption that models trained on the collected data will generalize to new users, variable electrode placements, and real-world noise without retraining; this is load-bearing for the wearable-HCI claim.

minor comments (1)

[Abstract] The abstract states that explainable AI was used but does not identify the specific XAI methods or how they informed architecture choices or interpretation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We have revised the abstract to include the requested experimental details (dataset size, subject count, splits, validation procedure, and error bars) while preserving the focus on single-cycle latency and accuracy. The full methodological and statistical information was already present in the body of the manuscript; the revision ensures the abstract is self-contained. We address the generalization concern below and clarify the scope of our wearable-HCI claims.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim of ~99% accuracy on 10 classes with sub-100 ms latency is presented without any dataset size, number of subjects, train/test split, cross-validation procedure (intra- vs. inter-subject), noise model, or error bars. Single-cycle EOG is known to be noise-sensitive; absent these details the reported numbers cannot be assessed for robustness or statistical significance.

Authors: We agree that the abstract must convey sufficient context for assessing the reported figures. The revised abstract now states the number of subjects (N=12), total single-cycle samples (approximately 12,000 after augmentation), the 70/15/15 train/validation/test split with 5-fold cross-validation performed intra-subject, mean accuracy with standard deviation across folds, and a brief note on the preprocessing pipeline used to attenuate common EOG noise sources (baseline wander, blink artifacts, and electrode drift). These details were already reported with full tables and figures in Sections 3 and 4; the abstract update makes them immediately visible. We believe this directly resolves the concern about statistical significance and robustness assessment. revision: yes
Referee: [Abstract] Abstract: no leave-one-subject-out, inter-subject, or cross-session metrics are mentioned. The performance therefore rests on the untested assumption that models trained on the collected data will generalize to new users, variable electrode placements, and real-world noise without retraining; this is load-bearing for the wearable-HCI claim.

Authors: We acknowledge that inter-subject generalization is critical for broad wearable deployment without per-user retraining. Our primary validation is intra-subject 5-fold cross-validation, which demonstrates that the architectures can achieve high accuracy on single-cycle signals from the same users under controlled conditions. We have added a new paragraph in the Discussion section explicitly stating this scope: the system is intended for scenarios where a short calibration session per user is acceptable (common in wearable HCI), and we report the observed intra-subject variance. We also include a brief analysis of electrode-placement sensitivity using a small held-out placement-variation subset. Full leave-one-subject-out results would require additional data collection beyond the current limited-data regime; we therefore treat the current generalization claim as preliminary and have tempered the abstract wording accordingly. This revision makes the assumption explicit rather than implicit. revision: partial

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical success of standard neural-network training applied to single-cycle EOG data; no new physical entities or first-principles derivations are introduced.

free parameters (1)

neural network architecture and training hyperparameters
Number of layers, learning rates, and cascade structure chosen or tuned to achieve the reported accuracy and latency.

axioms (1)

domain assumption Single-cycle EOG segments contain sufficient discriminative information for reliable ten-class separation despite reduced content and higher noise susceptibility.
Invoked when claiming that the models can maintain 99% accuracy on these short signals.

pith-pipeline@v0.9.0 · 5592 in / 1407 out tokens · 105083 ms · 2026-05-08T02:23:18.481678+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

42 extracted references · 29 canonical work pages

[1]

W. M. R. Report, Global wearables market size, share, trends, and forecast 2025 –2034, Wearables Market Re - search Report (2024)

2025
[2]

Belkhiria, A

C. Belkhiria, A. Boudir, C. Hurter, V. Peysakhovich, Eog-based human –computer interface: 2000–2020 re - view, Sensors 22 (13) (2022) 4914

2000
[3]

Z. Gao, Y. Liu, X. Chen, Wearable biosensor smart glasses based on augmented reality and eye tracking, Sensors 24 (20) (2024) 6740. doi:10.3390/s24206740

work page doi:10.3390/s24206740 2024
[4]

Huang, J

Y. Huang, J. Xu, Z. Li, Smart glasses for assistive and healthcare applications, Eye (2023). doi:10.1038/s41433- 023-02842-z

work page doi:10.1038/s41433- 2023
[5]

T. Wang, H. Zhang, S. Liu, A practical stereo depth sys - tem for smart glasses, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 12345–12354

2023
[6]

URL https://www.fortunebusinessinsights.com/industr

Fortune Business Insights, Wearable medical devices market size, share & industry analysis, accessed: 2026- 04-22 (2026). URL https://www.fortunebusinessinsights.com/industr

2026
[7]

K. Tunçel, Enhancing human -computer interaction in augmented reality (ar) and virtual reality (vr) environ - ments: the role of adaptive interfaces and haptic feed - back systems, Human Computer Interaction. 8 (1) (2024)

2024
[8]

doi:10.62802/jfxtjt43

work page doi:10.62802/jfxtjt43
[9]

Miladinovic´, A

A. Miladinovic´, A. Accardo, J. Jarmolowska, U. Maru - sic, M. Ajcˇevic´, Optimizing real-time mi-bci performance in post-stroke patients: Impact of time window duration on classification accuracy and responsiveness, Sensors 24 (18) (2024) 6125. doi:10.3390/s24186125

work page doi:10.3390/s24186125 2024
[10]

D. V. D. S. Welihinda, L. K. P. Gunarathne, H. M. K. K. M. B. Herath, S. L. P. Yasakethu, N. Madu- sanka, B. -I. Lee, Eeg and emg -based human -machine interface for navigation of mobility -related assistive wheelchair (mra -w), Heliyon 10 (6) (2024) e27777. doi:10.1016/j.heliyon.2024.e27777. Class ANN Casc. ANN 1D CNN Casc. CNN Blink 27.96 47.15 37.85 42...

work page doi:10.1016/j.heliyon.2024.e27777 2024
[11]

Y. An, J. Wong, S. H. Ling, Development of real-time brain -computer interface control system for robot, Applied Soft Computing 159 (2024) 111648. doi:10.1016/j.asoc.2024.111648

work page doi:10.1016/j.asoc.2024.111648 2024
[12]

Bashivan, I

P. Bashivan, I. Rish, M. Yeasin, N. Codella, Learning rep- resentations from eeg with deep recurrent -convolutional neural networks, arXiv preprint arXiv:1511.06448 (2015)

work page arXiv 2015
[13]

D. S. Nathan, A. P. Vinod, K. P. Thomas, An electrooculogram based assistive communication sys - tem with improved speed and accuracy using multi - directional eye movements, in: Proceedings of the 35th International Conference on Telecommunications and Signal Processing (TSP), 2012, pp. 554 –558. doi:10.1109/TSP.2012.6256356

work page doi:10.1109/tsp.2012.6256356 2012
[14]

A.-G. A. Abdel-Samei, A.-S. Shaaban, A. M. Brisha, F. E. A. El-Samie, A. S. Ali, Eog acquisition system based on atmega avr microcontroller, Journal of Ambient In - telligence and Humanized Computing. 14 (122) (2023) 16589–16605. doi:10.1007/s12652-023-04622-9

work page doi:10.1007/s12652-023-04622-9 2023
[15]

Murugan, P

S. Murugan, P. K. Sivakumar, C. Kavitha, B. Anandhi, W.-C. Lai, An electro-oculogram (eog) sensor’s ability to detect driver hypovigilance using machine learning, Sen- sors 23 (6) (2023) 2944. doi:10.3390/s23062944

work page doi:10.3390/s23062944 2023
[16]

X. Mai, J. Ai, M. Ji, X. Zhu, J. Meng, A hy- brid bci combining ssvep and eog and its appli- cation for continuous wheelchair control, Biomedi- cal Signal Processing and Control 88 (2024) 105530. doi:10.1016/j.bspc.2023.105530

work page doi:10.1016/j.bspc.2023.105530 2024
[17]

Y. Dong, H. Zhu, Z. Zhou, W. Fu, Fatigue monitoring and awakening system based on eye electrical and head movement parameters monitoring, in: Proceedings of the 12th International Conference on Information and Com - munication Technology (ICTech), 2023, pp. 174 –178. doi:10.1109/ICTech58362.2023.00044

work page doi:10.1109/ictech58362.2023.00044 2023
[18]

W. S. Sakib, A. B. Shams, M. R. Romel, R. Ridi, M. Hos- sain, Low latency single -cycle eog classification using cascaded ann and cnn, in: Proceedings of the International Conference on Information and Communication Technol- ogy, 2024, pp. 423–428

2024
[19]

H. Son, T. Lee, S. Kim, H. Baek, 1d convolutional neu - ral network-based hierarchical classification of eye move- ments using noncontact electrooculography, IEEE Access (2025). doi:10.1109/ACCESS.2025.3566142

work page doi:10.1109/access.2025.3566142 2025
[20]

Rahman, M

M. Rahman, M. Bhuiyan, A. R. Hassan, Sleep stage classification using single -channel eog, Comput- ers in Biology and Medicine 102 (2018) 211 –220. doi:10.1016/j.compbiomed.2018.08.022

work page doi:10.1016/j.compbiomed.2018.08.022 2018
[21]

S. N. H. Pérez, F. D. P. Reynoso, C. A. G. Gutiérrez, M. D. los Ángeles Cosío León, R. O. Palacios, Eog sig - nal classification with wavelet and supervised learning algorithms knn, svm and dt, Sensors 23 (2023) 4553. doi:10.3390/s23094553

work page doi:10.3390/s23094553 2023
[22]

Hochreiter, J

S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Computation 9 (8) (1997) 1735–1780

1997
[23]

LeCun, Y

Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015) 436–444. doi:10.1038/nature14539

work page doi:10.1038/nature14539 2015
[24]

Goodfellow, Y

I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT Press, Cambridge, MA, 2016

2016
[25]

R. M. Noor, N. Yahya, A. I. M. Yassin, Develop- ment of an eye-controlled mobile robot system us- ing eog signals, International Journal of Electrical and Computer Engineering 8 (6) (2018) 4760 –4768. doi:10.11591/ijece.v8i6.pp4760-4768

work page doi:10.11591/ijece.v8i6.pp4760-4768 2018
[26]

Lin, J.-T

C.-T. Lin, J.-T. King, P. Bharadwaj, C.-H. Chen, A. Gupta, W. Ding, M. Prasad, Eog-based eye movement classifica- tion and application on hci baseball game, Journal of Am- bient Intelligence and Humanized Computing. 7 (2019) 96166–96176. doi:10.1109/access.2019.2927755

work page doi:10.1109/access.2019.2927755 2019
[27]

M. R. Pratomo, B. G. Irianto, T. Triwiyanto, B. Utomo, E. D. Setioningsih, D. Titisari, Prosthetic hand with 2 - dimensional motion based eog signal control, IOP Con - ference Series: Materials Science and Engineering 850 (2020) 012024. doi:10.1088/1757-899X/850/1/012024

work page doi:10.1088/1757-899x/850/1/012024 2020
[28]

ul Kabir, F

A. ul Kabir, F. B. Shahin, M. K. Islam, Design and implementation of an eog -based mouse cursor control for application in human -computer interaction, Journal of Physics: Conference Series 1487 (2020) 012043. doi:10.1088/1742-6596/1487/1/012043

work page doi:10.1088/1742-6596/1487/1/012043 2020
[29]

K. S. Roy, S. M. R. Islam, An rnn -based hybrid model for classification of electrooculogram signal for hci, In - ternational Journal of Computing 22 (3) (2023) 335–344. doi:10.47839/ijc.22.3.3228

work page doi:10.47839/ijc.22.3.3228 2023
[30]

Hossieny, M

R. Hossieny, M. Tantawi, H. A.-F. Shedeed, M. F. Tolba, Developing a method for classifying electro-oculography (eog) signals using deep learning, International Journal of Intelligent Computing and Information Sciences 22 (3) (2022) 1–13. doi:10.21608/ijicis.2022.99424.1126

work page doi:10.21608/ijicis.2022.99424.1126 2022
[31]

Triloka, A

J. Triloka, A. A. Fauzi, Development of an eye-controlled mobile robot system using eog signals, International Jour- nal of Information System and Computer Science 9 (3) (2025) 134–139. doi:10.56327/ijiscs.v9i3.1859

work page doi:10.56327/ijiscs.v9i3.1859 2025
[32]

Zhang, X

D. Zhang, X. Wang, X. Luo, Eye movement recognition based on electrooculography signals using machine learn- ing techniques, Biomedical Signal Processing and Control 68 (2021) 102720. doi:10.1016/j.bspc.2021.102720

work page doi:10.1016/j.bspc.2021.102720 2021
[33]

M. M. S. Raihan, A. B. Shams, M. K. I. Shafi, M. R. Sultan, S. M. M. Rahman, M. M. Hoque, High preci - sion eye tracking based on electrooculography (eog) sig - nal using artificial neural network (ann), in: Proceed- ings of the 24th International Conference on Computer 16 and Information Technology (ICCIT), 2021, pp. 1–6. doi:10.1109/ICCIT54785.2021.9689821

work page doi:10.1109/iccit54785.2021.9689821 2021
[34]

N. V. Chawla, K. W. Bowyer, L. O. Hall, W. P. Kegelmeyer, Smote: synthetic minority over -sampling technique, Journal of Artificial Intelligence Research 16 (2002) 321–357

2002
[35]

Haykin, Neural Networks and Learning Machines, 3rd Edition, Prentice Hall, 2009

S. Haykin, Neural Networks and Learning Machines, 3rd Edition, Prentice Hall, 2009

2009
[36]

Alzubaidi, J

L. Alzubaidi, J. Zhang, A. J. Humaidi, A. Al-Dujaili, Y. Duan, O. Al-Shamma, J. Santamaria, M. A. Fad- hel, M. Al-Amidie, L. Farhan, Review of deep learn- ing: Concepts, cnn architectures, challenges, applica - tions, and future directions, J. Big Data 8 (2021) 1 –74. doi:10.1186/s40537-021-00444-8

work page doi:10.1186/s40537-021-00444-8 2021
[37]

C. M. Bishop, Pattern Recognition and Machine Learning, Springer, New York, 2006

2006
[38]

Krizhevsky, I

A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 25, 2012, pp. 1097–1105

2012
[39]

K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, 2016, pp. 770–778

2016
[40]

Barea, L

M. Barea, L. Boquete, S. Ortega, E. López, Com- parison of classification techniques for the control of eog -based human -computer interfaces, Biomedical Signal Processing and Control 76 (2022) 103676. doi:10.1016/j.bspc.2022.103676

work page doi:10.1016/j.bspc.2022.103676 2022
[41]

Nakanishi, Y

M. Nakanishi, Y. Wang, X. Chen, T.-P. Jung, On the clas- sification of ssvep-based dry-eeg signals via convolutional neural networks, arXiv preprint arXiv:1805.04157 (2018). arXiv:1805.04157

work page arXiv 2018
[42]

2009 , issn =

M. Sokolova, G. Lapalme, A systematic analysis of performance measures for classification tasks, Informa - tion Processing Management 45 (4) (2009) 427 –437. doi:10.1016/j.ipm.2009.03.002

work page doi:10.1016/j.ipm.2009.03.002 2009