pith. sign in

arxiv: 2606.09681 · v2 · pith:4QX5VHT7new · submitted 2026-06-08 · 💻 cs.CV

GenEyePose: Patient-Free, Knowledge-Based Saccadic Eye Movement Modeling for Digital Neurophysiologic Biomarker Development

Tianyu Lin , Jooyoung Ryu , Puvada Sreevarsha , Rahul Srinivasaragavan , Riya Satavlekar , Susan Kim , Nidhi Soley , Yujie Yan
show 7 more authors
Ishan Vatsaraj Carl Harris Aimon Rahman Vishal Patel Joseph Greenstein Casey Taylor Kemar E. Green
This is my paper

Pith reviewed 2026-06-27 17:17 UTC · model grok-4.3

classification 💻 cs.CV
keywords synthetic data generationsaccadic eye movementsdeep learning classifierdigital biomarkershypometriahypermetriavideo-oculographyneurophysiologic states
0
0 comments X

The pith

A fully synthetic pipeline generates saccadic eye movement data that trains classifiers to detect abnormal saccades on real clinical recordings with an AUROC of 0.76.

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

The paper introduces a patient-free method to create synthetic eye movement data based on knowledge of normal and abnormal saccade kinematics. This addresses the lack of datasets due to privacy concerns and enables training of AI models for detecting neurophysiologic biomarkers. A deep learning classifier trained only on this synthetic data distinguishes normal from abnormal saccades, such as hypometria and hypermetria, when tested on real-world clinical data. It achieves an AUROC of 0.76 and sensitivity of 0.71, indicating potential for applications in screening and brain abnormality localization without requiring real patient data for training.

Core claim

The central claim is that a fully synthetic, patient-free, multimodal eye movement generation pipeline can produce data that allows training a deep learning classifier to distinguish between normal and abnormal saccadic accuracies, and this classifier generalizes to real-world clinical data with an AUROC of 0.76 and sensitivity of 0.71.

What carries the argument

The knowledge-based synthetic generation pipeline that models saccadic eye movements, including normal and abnormal accuracies like hypometria and hypermetria, without using any patient data.

If this is right

  • Classifiers trained on synthetic data can achieve useful performance on unseen real clinical recordings for saccade abnormality detection.
  • The approach enables development of digital biomarkers for neurologic diseases without privacy barriers or scarce real datasets.
  • Such models have potential as screening tools in at-home and emergency room settings.
  • The method could support precise neuroanatomic localization of brain abnormalities.

Where Pith is reading between the lines

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

  • Extending the pipeline to other eye movement types could broaden its use in neurology.
  • Integration into portable devices might allow widespread at-home monitoring of neurophysiologic states.
  • Further validation on larger real datasets would strengthen the case for clinical deployment.
  • The synthetic approach might apply to modeling other physiologic signals where data is scarce.

Load-bearing premise

The knowledge-based rules for generating synthetic saccades accurately capture the key statistical and kinematic features of real abnormal eye movements.

What would settle it

If a new collection of real clinical eye movement recordings shows the synthetic-trained classifier performing no better than random guessing on distinguishing normal from abnormal saccades, the generalization claim would be falsified.

Figures

Figures reproduced from arXiv: 2606.09681 by Aimon Rahman, Carl Harris, Casey Taylor, Ishan Vatsaraj, Jooyoung Ryu, Joseph Greenstein, Kemar E. Green, Nidhi Soley, Puvada Sreevarsha, Rahul Srinivasaragavan, Riya Satavlekar, Susan Kim, Tianyu Lin, Vishal Patel, Yujie Yan.

Figure 1
Figure 1. Figure 1: GenEyePose framework. Clinical Practice: Pupils are extracted from real patient eye videos to generate waveforms for diagnosis. Proposed Pipeline: Synthetic eye movement videos are generated from synthetic waveforms to train deep classifiers. These movements are initiated and controlled by complex neural pathways in brain regions spanning the cortex, subcortex, brainstem, and cerebellum [11]. Saccadic metr… view at source ↗
Figure 2
Figure 2. Figure 2: The overview of the proposed eye movement video generation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Synthetic Waveform Validation. Main sequence analysis of our synthetic waveform shows high correlation with the amplitude-velocity curve of a previous study. 4.2 Synthetic Video Validation After using the waveforms to guide the movement of our pupil mask and subse￾quently our eye videos, we re-segmented the pupil mask from the eye videos to validate whether the videos correctly followed the movement in the… view at source ↗
Figure 4
Figure 4. Figure 4: Grad-CAM visualization on clinical data. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

Eye movements, including saccades, are widely regarded as highly sensitive and objective biomarkers of neurophysiologic states. Detecting saccadic signatures in neurologic diseases offers a rapid, portable alternative to brain imaging, avoiding access and cost barriers. Currently, there are no robust AI-enabled video-oculographic solutions (e.g., digital biomarkers) for screening, triaging, or localizing brain abnormalities due to privacy issues and scarce datasets. In this work, we propose the first fully synthetic, patient-free, multimodal eye movement generation pipeline for generalizable saccade analysis. Using this synthetic dataset, we trained a deep learning classifier to distinguish between normal and abnormal (hypometria and hypermetria) saccadic accuracies and evaluated its performance on real-world clinical data. The model achieved an AUROC of 0.76 and a sensitivity of 0.71, showing that the synthetic data has strong potential to generalize for clinical applications, including as a screening tool in at-home and emergency room settings or a tool for precise neuroanatomic localization.

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 introduces GenEyePose, a patient-free, knowledge-based synthetic pipeline for generating multimodal saccadic eye movement trajectories that model normal saccades as well as abnormal ones (hypometria and hypermetria). A deep-learning classifier is trained exclusively on the resulting synthetic dataset to distinguish normal from abnormal saccadic accuracy and is then evaluated on separate real-world clinical recordings, yielding an AUROC of 0.76 and sensitivity of 0.71.

Significance. If the synthetic generator is shown to reproduce the joint statistics of real abnormal saccades, the approach would directly address data scarcity and privacy barriers that currently limit video-oculographic biomarker development, potentially enabling scalable screening and neuroanatomic localization tools for at-home or emergency settings.

major comments (2)
  1. [Abstract] Abstract and Results: the headline generalization claim (AUROC 0.76, sensitivity 0.71 on real clinical data after training only on synthetic data) requires that the knowledge-based generator reproduces the joint distributions of saccadic accuracy, peak velocity, and latency for the hypometria and hypermetria classes; no Kolmogorov-Smirnov tests, Wasserstein distances, or overlaid histograms comparing synthetic abnormal trajectories to the real patient recordings used for evaluation are reported.
  2. [Methods] Methods: the description of the synthetic generation procedure supplies no quantitative validation against real distributions, no dataset sizes for the synthetic corpus, and no controls for distribution shift, leaving the central claim that the model learns pathology-specific signatures rather than generic video features unsupported by evidence.
minor comments (2)
  1. [Results] Clarify whether the reported AUROC and sensitivity are computed on a held-out real test set that is fully independent of any parameter tuning performed on synthetic data.
  2. [Abstract] The abstract states the pipeline is 'multimodal'; the manuscript should explicitly list the modalities generated and used by the classifier.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which identify key areas where additional validation would strengthen the central claims regarding the synthetic generator. We address each point below and will incorporate the requested analyses in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract and Results: the headline generalization claim (AUROC 0.76, sensitivity 0.71 on real clinical data after training only on synthetic data) requires that the knowledge-based generator reproduces the joint distributions of saccadic accuracy, peak velocity, and latency for the hypometria and hypermetria classes; no Kolmogorov-Smirnov tests, Wasserstein distances, or overlaid histograms comparing synthetic abnormal trajectories to the real patient recordings used for evaluation are reported.

    Authors: We agree that explicit statistical comparisons are needed to support the generalization claim. In the revised manuscript we will add Kolmogorov-Smirnov tests, Wasserstein distances, and overlaid histograms comparing the synthetic hypometria and hypermetria trajectories to the real clinical recordings for saccadic accuracy, peak velocity, and latency. revision: yes

  2. Referee: [Methods] Methods: the description of the synthetic generation procedure supplies no quantitative validation against real distributions, no dataset sizes for the synthetic corpus, and no controls for distribution shift, leaving the central claim that the model learns pathology-specific signatures rather than generic video features unsupported by evidence.

    Authors: We acknowledge these omissions. The revised Methods section will report the exact sizes of the synthetic corpus, include the quantitative distribution comparisons noted above, and add an explicit discussion of the knowledge-based parameter ranges used to control distribution shift and ensure pathology-specific signatures are modeled rather than generic video features. revision: yes

Circularity Check

0 steps flagged

No significant circularity; evaluation on held-out real data is independent of synthetic generator

full rationale

The paper's central result is an AUROC of 0.76 obtained by training exclusively on synthetic saccades and testing on separate real clinical recordings. No equations, fitted parameters, or self-citations are shown that would make this metric equivalent to the generator's inputs by construction. The knowledge-based synthesis step is presented as an external modeling choice whose fidelity is tested rather than assumed tautologically. This is a standard train-on-synthetic/test-on-real setup with no load-bearing self-referential reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no information is provided on free parameters, axioms, or invented entities used in the knowledge-based modeling.

pith-pipeline@v0.9.1-grok · 5782 in / 1187 out tokens · 18854 ms · 2026-06-27T17:17:14.607008+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

31 extracted references · 8 canonical work pages

  1. [1]

    Experimental neurology48(1), 107–122 (1975)

    Bahill, A.T., Clark, M.R., Stark, L.: Dynamic overshoot in saccadic eye movements is caused by neurological control signal reversals. Experimental neurology48(1), 107–122 (1975)

    1975

  2. [2]

    Computer methods and programs in biomedicine 84(2-3), 174–187 (2006)

    De Santis, A., Iacoviello, D.: Optimal segmentation of pupillometric images for es- timating pupil shape parameters. Computer methods and programs in biomedicine 84(2-3), 174–187 (2006)

    2006

  3. [3]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6824–6835 (2021)

    2021

  4. [4]

    2021 IEEE Interna- tional Symposium on Mixed and Augmented Reality (ISMAR) pp

    Fuhl, W., Kasneci, G., Kasneci, E.: Teyed: Over 20 million real-world eye im- ages with pupil, eyelid, and iris 2d and 3d segmentations, 2d and 3d land- marks, 3d eyeball, gaze vector, and eye movement types. 2021 IEEE Interna- tional Symposium on Mixed and Augmented Reality (ISMAR) pp. 367–375 (2021), https://api.semanticscholar.org/CorpusID:231786490

    2021

  5. [5]

    Behavior Research Methods pp

    Gibaldi, A., Sabatini, S.P.: The saccade main sequence revised: A fast and repeat- able tool for oculomotor analysis. Behavior Research Methods pp. 167–187 (2020), https://link.springer.com/article/10.3758/s13428-020-01388-2

    work page doi:10.3758/s13428-020-01388-2 2020

  6. [6]

    Big Data and Cognitive Computing8(12), 198 (2024)

    Graham, L., Vitorio, R., Walker, R., Barry, G., Godfrey, A., Morris, R., Stuart, S.: Digital eye-movement outcomes (demos) as biomarkers for neurological conditions: A narrative review. Big Data and Cognitive Computing8(12), 198 (2024)

    2024

  7. [7]

    In: 2011 International Joint Conference on Biometrics (IJCB)

    Holland, C., Komogortsev, O.V.: Biometric identification via eye movement scan- paths in reading. In: 2011 International Joint Conference on Biometrics (IJCB). pp. 1–8 (2011). https://doi.org/10.1109/IJCB.2011.6117536

    work page doi:10.1109/ijcb.2011.6117536 2011

  8. [8]

    arXiv preprint arXiv:2409.01240 (2024)

    Jiao, C., Zhang, G., Hu, Z., Bulling, A.: Diffeyesyn: Diffusion-based user-specific eye movement synthesis. arXiv preprint arXiv:2409.01240 (2024)

    work page arXiv 2024

  9. [9]

    JAMA Otolaryngology–Head & Neck Surgery148(5), 474– 475 (2022)

    Kattah, J.C., Newman-Toker, D.E.: Video-oculography to guide neuroimaging for dizziness and vertigo. JAMA Otolaryngology–Head & Neck Surgery148(5), 474– 475 (2022)

    2022

  10. [10]

    Muscle & nerve64(3), 328–335 (2021)

    Kocak, G.S., Tütüncü, M., Adatepe, N.U., Yerlikaya, B.D., Kara, E., Atas, A., Yener, M., Oren, M.M.: A novel diagnostic method for myasthenia gravis. Muscle & nerve64(3), 328–335 (2021)

    2021

  11. [11]

    Oxford university press (2015)

    Leigh, R.J., Zee, D.S.: The neurology of eye movements. Oxford university press (2015)

    2015

  12. [12]

    In: Proceedings of the 18th ACM international conference on Multimedia

    Marcel, S., Rodriguez, Y.: Torchvision the machine-vision package of torch. In: Proceedings of the 18th ACM international conference on Multimedia. pp. 1485– 1488 (2010)

    2010

  13. [13]

    Mei, K., Patel, V.M.: Vidm: Video implicit diffusion models (2022)

    2022

  14. [14]

    bioRxiv pp

    Mukunda, K.N., Ye, T., Luo, Y., Zoitou, A., Kwon, K.E., Singh, R., Woo, J., Sivakumar, N., Greenstein, J.L., Taylor, C.O., et al.: Deep learning detection of subtle torsional eye movements: Preliminary results. bioRxiv pp. 2024–05 (2024)

    2024

  15. [15]

    Controlnext: Powerful and efficient control for image and video generation.arXiv preprint arXiv:2408.06070,

    Peng, B., Wang, J., Zhang, Y., Li, W., Yang, M.C., Jia, J.: Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070 (2024)

    work page arXiv 2024

  16. [16]

    Sensors (Basel)23(4) (February 2023)

    Przybyszewski, A.W., Śledzianowski, A., Chudzik, A., Szlufik, S., Koziorowski, D.: Machine learning and eye movements give insights into neurodegenerative disease mechanisms. Sensors (Basel)23(4) (February 2023)

    2023

  17. [17]

    bioRxiv pp

    Rahman, A., Bachina, P., Patel, V., Green, K.E.: Generative artificial intelligence for secure and scalable multimodal eye movement datasets. bioRxiv pp. 2025–04 (2025) 10 T. Lin et al

    2025

  18. [18]

    In: Ali, S., Hogg, D.C., Peckham, M

    Rahman, A., Green, K.E., Patel, V.M.: Genvog: A diffusion probabilistic frame- work for patient-independent pose-guided nystagmus video-oculography (vog) generation. In: Ali, S., Hogg, D.C., Peckham, M. (eds.) Medical Image Un- derstanding and Analysis. Lecture Notes in Computer Science, vol. 15918, pp. 306–316. Springer, Cham (2026). https://doi.org/10.1...

    work page doi:10.1007/978-3-031-98694-9_22 2026

  19. [19]

    Medical Imaging with Deep Learning (2026), https://openreview.net/forum?id= OGjV2u9p8n, mIDL 2026 Poster

    Rahman, A., Green, K.E., Patel, V.M.: Genvog-dit: A transformer-based dif- fusion model for pose-driven, patient-agnostic nystagmus vog video generation. Medical Imaging with Deep Learning (2026), https://openreview.net/forum?id= OGjV2u9p8n, mIDL 2026 Poster

    2026

  20. [20]

    Brain130(1), 10–35 (11 2006)

    Ramat, S., Leigh, R.J., Zee, D.S., Optican, L.M.: What clinical disorders tell us about the neural control of saccadic eye movements. Brain130(1), 10–35 (11 2006). https://doi.org/10.1093/brain/awl309, https://doi.org/10.1093/brain/awl309

    work page doi:10.1093/brain/awl309 2006

  21. [21]

    Computer Vision and Image Understanding170, 40–50 (2018)

    Santini, T., Fuhl, W., Kasneci, E.: Pure: Robust pupil detection for real-time per- vasive eye tracking. Computer Vision and Image Understanding170, 40–50 (2018)

    2018

  22. [22]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

    Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad- cam: Visual explanations from deep networks via gradient-based localization. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 618–626 (2017). https://doi.org/10.1109/ICCV.2017.74

    work page doi:10.1109/iccv.2017.74 2017

  23. [23]

    bioRxiv (2023)

    Steinhart, B., Brooks-Russell, A., Kosnett, M.J., Subramanian, P.S., Wrobel, J.: A video segmentation pipeline for assessing changes in pupil response to light after cannabis consumption. bioRxiv (2023)

    2023

  24. [24]

    The Lancet Neurology23(4), 344–381 (2024)

    Steinmetz, J.D., Seeher, K.M., Schiess, N., Nichols, E., Cao, B., Servili, C., Cav- allera, V., Cousin, E., Hagins, H., Moberg, M.E., et al.: Global, regional, and national burden of disorders affecting the nervous system, 1990–2021: a systematic analysis for the global burden of disease study 2021. The Lancet Neurology23(4), 344–381 (2024)

    1990

  25. [25]

    PLOS One (2018)

    Susanne, H., Liesenfeld, M., Schmidtmann, I., Ashayer, S., Pitz, S.: Age dependent normative data of vertical and horizontal reflexive saccades. PLOS One (2018)

    2018

  26. [26]

    In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition

    Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 6450–6459 (2018)

    2018

  27. [27]

    Frontiers in Neurology13, 963968 (2022)

    Wagle, N., Morkos, J., Liu, J., Reith, H., Greenstein, J., Gong, K., Gangan, I., Pakhomov, D., Hira, S., Komogortsev, O.V., et al.: aEYE: A deep learning system for video nystagmus detection. Frontiers in Neurology13, 963968 (2022)

    2022

  28. [28]

    Wang, C., Bai, Y., Tsang, A., Bian, Y., Gou, Y., Lin, Y.X., Zhao, M., Wei, T.Y., Desman, J.M., Taylor, C.O., et al.: Deep learning model for static ocular torsion detectionusingsyntheticallygeneratedfundusimages.TranslationalVisionScience & Technology12(1), 17–17 (2023)

    2023

  29. [29]

    medRxiv pp

    Wei, H., Bosley, J., Kuwera, E., Kazanzides, P., Green, K.E.: Teleautohints: A vir- tual or augmented reality (vr/ar) system for automated tele-neurologic evaluation of acute vertigo. medRxiv pp. 2025–03 (2025)

    2025

  30. [30]

    Journal of Neuroscience Methods324, 108307 (2019)

    Yiu, Y.H., Aboulatta, M., Raiser, T., Ophey, L., Flanagin, V.L., zu Eulenburg, P., Ahmadi, S.A.: Deepvog: Open-source pupil segmentation and gaze estimation in neuroscience using deep learning. Journal of Neuroscience Methods324, 108307 (2019). https://doi.org/https://doi.org/10.1016/j.jneumeth.2019.05.016

    work page doi:10.1016/j.jneumeth.2019.05.016 2019

  31. [31]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836–3847 (2023)

    2023