Integrating Eye Tracking and Inertial Sensing for Enhanced Freezing of Gait Detection in Parkinson's Disease

Integrating Eye Tracking and Inertial Sensing for Enhanced Freezing of Gait Detection in Parkinson's Disease | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL European Journal of Neuroscience This is a preprint and has not been peer reviewed. Data may be preliminary. 23 August 2025 V1 Latest version Share on Integrating Eye Tracking and Inertial Sensing for Enhanced Freezing of Gait Detection in Parkinson's Disease Authors : Christopher Pulliam 0000-0002-4623-6022 [email protected] , Jinxin Chen , and James Liao 0000-0001-6605-9855 Authors Info & Affiliations https://doi.org/10.22541/au.175597775.50778140/v1 Published European Journal of Neuroscience Version of record Peer review timeline 328 views 229 downloads Contents Abstract Supplementary Material Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Freezing of gait, a disabling symptom in Parkinson’s disease, presents a major challenge for wearable classification algorithms that struggle to distinguish pathological freezes from voluntary stops. To address this ambiguity, we evaluated whether incorporating eye-gaze kinematics could improve classification accuracy compared to using ankle-mounted inertial measurement units (IMUs) alone. We analyzed data from 10 participants performing standardized walking tasks and compared two deep learning classifiers differing only in their inputs: an IMU‑only model (bilateral ankle accelerometer and gyroscope) and an IMU+Gaze model that fuses IMU channels with 3D fixation‑velocity from the headset eye tracker. With subject‑independent five‑fold cross‑validation, the IMU+Gaze model improved macro‑averaged performance – precision 0.758 vs 0.675, recall 0.867 vs 0.701, and F1 0.801 vs 0.683 – relative to IMU‑only. Gains were largest for intentional standing (F1 0.679 vs 0.390), reducing the proportion of standing windows misclassified as freezing from 63.6% (14/22) to 9.1% (2/22). These findings show that gaze kinematics complement ankle kinematics for disambiguating voluntary stopping from FOG and potentially strengthen automated monitoring, clinician‑facing assessment, and patient‑facing assistive technologies. Title Integrating Eye Tracking and Inertial Sensing for Enhanced Freezing of Gait Detection in Parkinson’s Disease Running Title Eye Tracking and IMU Sensing for FOG Detection Authors Christopher L Pulliam, PhD 1 (ORCID: 0000-0002-4623-6022; e-mail: [email protected] ), Jinxin Chen, MS 1,2,* (ORCID: 0009-0000-0440-6460), James Y Liao, MD, PhD 2 (ORCID: 0000-0001-6605-9855) Affiliations Department of Biomedical Engineering, Case Western Reserve University, Cleveland, Ohio, United States of America Center for Neurological Restoration, Neurological Institute, Cleveland Clinic, Cleveland, Ohio, United States of America Correspondence Christopher L Pulliam ( [email protected] ) Acknowledgments This is a secondary analysis of data collected under a prior study supported by the American Parkinson’s Disease Association. The funder had no role in the current analysis, interpretation, or manuscript preparation. Relevant Financial Disclosures / Conflict of Interest JYL is an inventor on intellectual property related to the technology described in this manuscript. This intellectual property has been licensed to Strolll Limited; JYL is eligible to receive future royalties. The remaining authors report no competing interests. Words: 1826 Figures: 1 Tables: 1 *JC is presently with The University of Alabama at Birmingham, Birmingham, Alabama, United States of America Abstract Freezing of gait, a disabling symptom in Parkinson’s disease, presents a major challenge for wearable classification algorithms that struggle to distinguish pathological freezes from voluntary stops. To address this ambiguity, we evaluated whether incorporating eye-gaze kinematics could improve classification accuracy compared to using ankle-mounted inertial measurement units alone. We analyzed data from 10 participants performing standardized walking tasks and compared two deep learning classifiers differing only in their inputs: an IMU‑only model (bilateral ankle accelerometer and gyroscope) and an IMU+Gaze model that fuses IMU channels with 3D fixation‑velocity from the headset eye tracker. With subject‑independent five‑fold cross‑validation, the IMU+Gaze model improved macro‑averaged performance – precision 0.758 vs 0.675, recall 0.867 vs 0.701, and F1 0.801 vs 0.683 – relative to IMU‑only. Gains were largest for intentional standing (F1 0.679 vs 0.390), reducing the proportion of standing windows misclassified as freezing from 63.6% (14/22) to 9.1% (2/22). These findings show that gaze kinematics complement ankle kinematics for disambiguating voluntary stopping from FOG and potentially strengthen automated monitoring, clinician‑facing assessment, and patient‑facing assistive technologies. Key Words Parkinson’s disease, freezing, inertial measurement unit, machine learning, eye tracking, digital biomarkers Abbreviations AR, augmented reality; BiLSTM, bidirectional long short-term memory; FOG, freezing of gait; IMU, inertial measurement unit; ML2, Magic Leap 2; PD, Parkinson’s disease; ReLU, rectified linear unit Introduction Freezing of gait (FOG) is a common, burdensome symptom of Parkinson’s disease (PD), characterized by sudden, brief inability to move the feet despite the intention to walk and often leading to falls and reduced mobility.(1–3) FOG episodes present as complete cessation of stepping (akinesia), trembling in place, or festinating steps.(4,5) External cueing with sensory stimuli (auditory, vibrotactile, or visual) can alleviate FOG for many individuals(6–15) and there is growing interest in on‑demand cueing that triggers only when freezing is detected to personalize therapy and reduce attentional burden.(16) Advancing these approaches requires more accurate detection algorithms that support not only potential on‑demand assistive cues but also reliable monitoring of FOG burden, objective clinical assessment, and clinical trial endpoints. Current wearable detectors based on inertial measurement units (IMUs) show promise,(17,18) but a key limitation is distinguishing FOG from voluntary stops.(19) The kinematics of akinetic freezes, in particular, can resemble intentional halts. Multimodal strategies have therefore explored physiological signals (e.g., heart rate variability) to augment motion data.(19–21) In parallel, evidence links oculomotor impairments to FOG,(22–25) and changes in oculomotor control (e.g., saccadic execution and fixation stability) may precede freezing, especially during turning tasks (26). With wearable, head‑mounted eye tracking now able to continuously record gaze during ambulation, concurrent capture of gaze and body motion is feasible in standardized assessments and everyday settings. Here, we evaluate whether incorporating eye‑gaze kinematics improves automated FOG detection. We compare an IMU‑only model with a multimodal IMU+Gaze model that fuses ankle IMU channels with 3D fixation‑velocity. We hypothesized that gaze would increase classification performance, particularly for differentiating voluntary standing from freezing, thereby improving utility for monitoring, assessment, and assistive applications. Materials and Methods 2.1 Source Dataset We conducted a secondary analysis of data from the ELIMINATE FoG clinical trial, which investigated augmented‑reality cueing in 36 individuals with PD.(15) The present study focused on 10 participants with high‑quality, synchronized eye‑tracking. Only the no‑cue control condition was analyzed to avoid cue‑induced alterations of gaze or gait during detection. 2.2 Data Acquisition and Labeling Body motion and eye-tracking data were captured simultaneously. Six Opal inertial measurement units (APDM Wearable Technologies) were worn on the ankles, wrists, and waist, and an ML2 augmented‑reality headset (Magic Leap, Inc.) provided integrated eye tracking. IMUs sampled triaxial accelerometer and gyroscope signals at 128 Hz; the headset recorded gaze position and fixation points at 60 Hz, interpolated to 128 Hz for temporal alignment with the IMU data streams. Although multiple IMUs were worn, only the two ankle‑mounted IMUs were used as model inputs. All trials were video‑recorded and annotated by trained raters to produce frame‑level labels of four motor states: sitting, standing, walking, or freezing. Classification targeted walking, freezing, and standing; sitting was excluded. Voluntary standing predominantly occurred during structured pauses in the protocol (e.g., trial start/end or between walking segments). 2.3 Data Pre-Processing Raw sensor data were segmented into three-second analysis windows with a 50% overlap (1.5-second step size). Analyses were performed offline on fixed windows, emphasizing steady‑state segments; onset latency was not estimated. Windows were retained only if they contained a single, consistent label (walking, freezing, or standing) and the ML2 headset’s “fixation confidence” (i.e., its internal estimate of gaze tracking quality on a 0-1 scale) exceeded zero throughout the window. We extracted 15 time-series signals per window as model inputs: 12 signals from ankle IMUs (3-axis accelerometer and 3-axis gyroscope from each ankle) and 3 signals representing 3D fixation‑velocity (first difference of fixation position), chosen to reduce confounding from fixed visual targets in the environment. 2.4 Classification Model Architecture and Training We developed and compared two deep learning models to isolate the contribution of gaze kinematics: (1) an IMU-Only Model using the 12 ankle IMU signals, and (2) a Multimodal Model using all 15 signals (12 IMU + 3 eye-fixation velocity). Both models were built on a bidirectional long short-term memory (BiLSTM) network architecture designed to capture temporal patterns in the sensor data.(27) The network architecture processed time-domain signals through an input layer, followed by a BiLSTM layer (50 units) with return sequences enabled and 40% dropout for regularization. A second BiLSTM layer (25 units) without return sequences processed this output, followed by another 40% dropout layer. The resulting features passed through a fully connected dense layer (10 units) with ReLU activation. Finally, a softmax output layer with 3 units produced class probabilities for walking, freezing, or standing. Hyperparameters were pre‑specified and not tuned on the test folds. All inputs were z‑score normalized, with normalization parameters estimated on training folds and applied to held‑out folds to prevent leakage. Models were trained for 75 epochs with the Adam optimizer and class‑weighted cross‑entropy to mitigate class imbalance. Analyses were conducted in MATLAB R2024b using the Deep Learning and Statistics and Machine Learning Toolboxes. 2.5 Evaluation and Statistical Analysis Performance was assessed with five‑fold, subject‑independent cross‑validation, withholding two participants per fold to estimate generalization to unseen individuals. We report fold‑wise means for overall accuracy and for precision, recall, and F1 at the class level, as well as the macro‑averaged (unweighted) F1 across classes. We also present aggregate confusion matrices to illustrate error modes (e.g., misclassification of standing as freezing). Results A total of 1,046 three-second windows met all quality criteria and were analyzed. Of these, 874 were labeled walking, 150 freezing, and 22 standing. Table 1 presents cross-validated precision, recall, and F1 scores for each activity class and the unweighted macro averages for both models. Adding gaze kinematics improved every macro metric: precision increased from 0.675 to 0.758, recall from 0.701 to 0.867, and F1 from 0.683 to 0.801. Class‑level results show a similar pattern. Performance for walking remained high, with negligible change in F1 (from 0.953 to 0.955). For freezing, recall was comparable between models (0.793 vs 0.807), but precision rose from 0.627 to 0.748, producing a higher F1 (from 0.706 to 0.770). The largest gain occurred for standing: precision increased from 0.421 to 0.559, recall from 0.364 to 0.864, and F1 nearly doubled from 0.390 to 0.679. Figure 1 shows the underlying confusion matrices for both models. In the IMU-only condition (right), 14 out of 22 standing windows were (63.6%) incorrectly labeled as freezing. With gaze information (left), this error was reduced to 2 out of 22 windows (9.1%), indicating improved discrimination between intentional stillness and pathological freezing. Gaze data also reduced confusion between walking and freezing: walking windows incorrectly classified as freezing dropped from 58 to 38, while freezing windows misclassified as standing decreased from 9 to 4. These improvements increased overall accuracy from 90.2% to 92.1% without compromising detection of true freezing episodes, as evidenced by stable recall for the freezing class. Discussion This study demonstrates that incorporating eye-gaze kinematics with inertial sensor data provides a substantial advantage for the automated detection of FOG, primarily by resolving the long-standing ambiguity between pathological freezing and intentional stops. Our multimodal model significantly outperformed an IMU-only model, achieving higher macro-averaged precision, recall, and F1-scores. The improvement was most dramatic for the classification of standing, where the F1-score nearly doubled, confirming our hypothesis that gaze provides critical, non-redundant information about a user’s motor state. The improved performance appears to stem from the availability of gaze information that helps distinguish voluntary motor arrest from pathological freezing. Existing literature provides a strong neurophysiological basis for this distinction. It is known, for instance, that people with PD who experience FOG exhibit greater oculomotor impairment (i.e., longer saccade latencies and increased movement variability) compared to those without FOG.(23) Furthermore, such oculomotor deficits have been tied to underlying alterations in the shared neural networks that control both gait and gaze.(22) Other studies have shown that just before a freeze, gaze can become unstable and deviate from the intended path of movement.(26) Taken together, this body of evidence suggests that the period just before and during a FOG episode may exhibit a distinct and detectable pattern of oculomotor dysfunction. It is therefore plausible that our deep learning model learned to identify temporal signatures in the fixation velocity data that correspond to these differing motor states, thereby improving its classification accuracy. Previous efforts to reduce FOG misclassification have focused on combining inertial signals with physiological measures such as heart-rate variability, skin conductance, or surface EMG. (19–21,28,29) Although those studies reported modest sensitivity gains, voluntary halts can still be labeled as FOG, particularly when akinesia lacked the characteristic burst in ankle freezing band power.(19) Our findings demonstrate that gaze kinematics – captured unobtrusively with a commercial AR headset – provide non-redundant information that substantially improves specificity without sacrificing sensitivity. From a practical standpoint, the primary near-term implication of adding eye gaze is improved specificity for monitoring and assessment (e.g., reducing false positives when quantifying FOG burden over repeated clinic visits or during structured home assessments). Greater discrimination between intentional stops and FOG can also improve phenotyping and evaluation of treatment response in trials. Assistive applications remain plausible; however, because the present bidirectional architecture leverages future context within each three second window and is therefore non-causal, realizing low-latency closed-loop cueing will require causal models, shorter decision windows, and on-device optimization. Several additional limitations warrant consideration. First, our findings are constrained by the study’s design: the analysis relied on a small cohort performing structured tasks in a controlled environment and included only 22 standing windows. These factors limit the statistical power for the standing class and mean that performance in more ecologically valid, everyday settings remains unknown. Second, our methodology itself imposes constraints. Relying on one architecture also limits generalizability, as alternative sequence models (causal LSTMs, temporal convolutional neural networks, lightweight transformers), different fusion strategies (late- or attention-based fusion), and shorter windows could alter class-specific performance, particularly for standing. Moreover, without feature-attribution or ablation analyses we cannot localize which gaze features drive improvement. Finally, the eye tracker’s 60 Hz sampling is insufficient for reliable classification of discrete oculomotor events such as saccades (30), so although the model shows that gaze contributes, we cannot identify the specific oculomotor mechanisms underlying the observed gains. Future work should directly address these gaps. A next step is to validate the added value of gaze kinematics in larger, more diverse cohorts during unstructured activities of daily living. Such studies are essential to clarify generalizability and determine whether the observed reduction in false positives translates into tangible benefits for closed-loop FOG management. Additionally, mechanistic studies using research-grade, high-frequency eye-trackers are needed to dissect the specific contributions of microsaccades, saccadic velocity, and fixation stability, thereby providing deeper insight into the neurophysiological link between oculomotor control and gait freezing. Conclusion In this study, we demonstrated that incorporating eye-gaze kinematics with inertial sensor data significantly improves the automated detection of FOG. The primary benefit of this multimodal approach was the enhanced ability to distinguish intentional standing from pathological freezing, a weakness of algorithms relying on body-worn sensors alone. This improved specificity, achieved without compromising detection of true FOG episodes, supports gaze‑augmented models as digital biomarkers for monitoring and objective assessment. These models may also inform future assistive systems, contingent on validation in free‑living settings and causal, low‑latency implementation. Data Availability Statement This study is a secondary analysis of data from a separate study led by author James Y Liao. De-identified data may be released through application to the Cleveland Clinic Institutional Review Board ( [email protected] ). Analysis code will be provided by the corresponding author upon reasonable request. References 1. Giladi N, Treves TA, Simon ES, Shabtai H, Orlov Y, Kandinov B, et al. Freezing of gait in patients with advanced Parkinson’s disease. J Neural Transm (Vienna). 2001;108(1):53–61. 2. Bloem BR, Hausdorff JM, Visser JE, Giladi N. Falls and freezing of gait in Parkinson’s disease: a review of two interconnected, episodic phenomena. Mov Disord. 2004 Aug;19(8):871–84. 3. Nonnekes J, Snijders AH, Nutt JG, Deuschl G, Giladi N, Bloem BR. Freezing of gait: a practical approach to management. The Lancet Neurology. 2015 Jul 1;14(7):768–78. 4. Schaafsma JD, Balash Y, Gurevich T, Bartels AL, Hausdorff JM, Giladi N. Characterization of freezing of gait subtypes and the response of each to levodopa in Parkinson’s disease. European Journal of Neurology. 2003;10(4):391–8. 5. Ehgoetz Martens KA, Shine JM, Walton CC, Georgiades MJ, Gilat M, Hall JM, et al. Evidence for subtypes of freezing of gait in Parkinson’s disease. Movement Disorders. 2018;33(7):1174–8. 6. Ginis P, Nackaerts E, Nieuwboer A, Heremans E. Cueing for people with Parkinson’s disease with freezing of gait: A narrative review of the state-of-the-art and novel perspectives. Ann Phys Rehabil Med. 2018 Nov;61(6):407–13. 7. Cosentino C, Putzolu M, Mezzarobba S, Cecchella M, Innocenti T, Bonassi G, et al. One cue does not fit all: A systematic review with meta-analysis of the effectiveness of cueing on freezing of gait in Parkinson’s disease. Neurosci Biobehav Rev. 2023 Jul;150:105189. 8. Nonnekes J, Nieuwboer A. Towards Personalized Rehabilitation for Gait Impairments in Parkinson’s Disease. J Parkinsons Dis. 2018;8(s1):S101–6. 9. Mancini M, Smulders K, Harker G, Stuart S, Nutt JG. Assessment of the ability of open- and closed-loop cueing to improve turning and freezing in people with Parkinson’s disease. Sci Rep. 2018 Aug 24;8(1):12773. 10. McCandless PJ, Evans BJ, Janssen J, Selfe J, Churchill A, Richards J. Effect of three cueing devices for people with Parkinson’s disease with gait initiation difficulties. Gait Posture. 2016 Feb;44:7–11. 11. Donovan S, Lim C, Diaz N, Browner N, Rose P, Sudarsky LR, et al. Laserlight cues for gait freezing in Parkinson’s disease: an open-label study. Parkinsonism Relat Disord. 2011 May;17(4):240–5. 12. Chan HL, Chen RS, Kuo CC, Chen YT, Liaw JW, Liao GS, et al. Laser-light cueing shoes with integrated foot pressure and inertial sensing for investigating the impact of visual cueing on gait characteristics in Parkinson’s disease individuals. Front Bioeng Biotechnol. 2024;12:1334403. 13. Salari M, Rashedi R, Momenilandi F, Shirin MB, Etemadifar M. The effects of laser cane cues on the freezing of gait of Parkinson’s disease patients: can increasing the laser light beam width play a role? Acta Neurol Belg. 2024 Oct;124(5):1647–53. 14. Geerse DJ, Coolen B, van Hilten JJ, Roerdink M. Holocue: A Wearable Holographic Cueing Application for Alleviating Freezing of Gait in Parkinson’s Disease. Front Neurol. 2022 Jan 10;12:628388. 15. Baugher B, Kaya R, Sonneborn C, Baker KB, Fernandez HH, Szewczyk N, et al. Effects of augmented reality cueing strategies on freezing of gait: The ELIMINATE FoG trial. Clin Park Relat Disord. 2025;12:100332. 16. Zoetewei D, Herman T, Ginis P, Palmerini L, Brozgol M, Thumm PC, et al. On-Demand Cueing for Freezing of Gait in Parkinson’s Disease: A Randomized Controlled Trial. Mov Disord. 2024 May;39(5):876–86. 17. Mancini M, McKay JL, Cockx H, D’Cruz N, Esper CD, Filtjens B, et al. Technology for measuring freezing of gait: Current state of the art and recommendations. J Parkinsons Dis. 2025 Feb;15(1):19–40. 18. Salomon A, Gazit E, Ginis P, Urazalinov B, Takoi H, Yamaguchi T, et al. A machine learning contest enhances automated freezing of gait detection and reveals time-of-day effects. Nat Commun. 2024 Jun 6;15(1):4853. 19. Cockx H, Nonnekes J, Bloem BR, van Wezel R, Cameron I, Wang Y. Dealing with the heterogeneous presentations of freezing of gait: how reliable are the freezing index and heart rate for freezing detection? J Neuroeng Rehabil. 2023 Apr 27;20(1):53. 20. Maidan I, Plotnik M, Mirelman A, A.Weiss, Giladi N, Hausdorff JM. Heart rate changes during freezing of gait in patients with Parkinson’s disease. Mov Disord. 2010 Oct 30;25(14):2346–54. 21. Economou K, Quek D, MacDougall H, Lewis SJG, Ehgoetz Martens KA. Heart Rate Changes Prior to Freezing of Gait Episodes Are Related to Anxiety. J Parkinsons Dis. 2021;11(1):271–82. 22. Gallea C, Wicki B, Ewenczyk C, Rivaud-Péchoux S, Yahia-Cherif L, Pouget P, et al. Antisaccade, a predictive marker for freezing of gait in Parkinson’s disease and gait/gaze network connectivity. Brain. 2021 Mar 3;144(2):504–14. 23. Nemanich ST, Earhart GM. Freezing of gait is associated with increased saccade latency and variability in Parkinson’s disease. Clin Neurophysiol. 2016 Jun;127(6):2394–401. 24. Wu L, Wang Q, Zhao L, Jiang CY, Xu Q, Wu SC, et al. Clinical and Oculomotor Correlates With Freezing of Gait in a Chinese Cohort of Parkinson’s Disease Patients. Front Aging Neurosci. 2020;12:237. 25. Zhou MX, Wang Q, Lin Y, Xu Q, Wu L, Chen YJ, et al. Oculomotor impairments in de novo Parkinson’s disease. Front Aging Neurosci. 2022;14:985679. 26. Wang Y, Wanrooij MM van, Emaus R, Nonneke J, Cohen MX, Cockx H, et al. Relationship between eye movements and freezing of gait during turning in individuals with Parkinson’s disease [Internet]. medRxiv; 2021 [cited 2025 Jun 13]. p. 2021.09.06.21263180. Available from: https://www.medrxiv.org/content/10.1101/2021.09.06.21263180v127. Sundermeyer M, Ney H, Schlüter R. From Feedforward to Recurrent LSTM Neural Networks for Language Modeling. IEEE/ACM Transactions on Audio, Speech, and Language Processing. 2015 Mar;23(3):517–29. 28. Moore A, Li J, Contag CH, Currano LJ, Pyles CO, Hinkle DA, et al. Wearable Surface Electromyography System to Predict Freeze of Gait in Parkinson’s Disease Patients. Sensors. 2024 Jan;24(23):7853. 29. Zhang W, Yang Z, Li H, Huang D, Wang L, Wei Y, et al. Multimodal Data for the Detection of Freezing of Gait in Parkinson’s Disease. Sci Data. 2022 Oct 7;9(1):606. 30. Leube A, Rifai K, Rifai K. Sampling rate influences saccade detection in mobile eye tracking of a reading task. J Eye Mov Res. 10(3):10.16910/jemr.10.3.3. (A | B ) & Recall (A | B ) F1 Score (A | B ) Walking 0.967 | 0.976 0.944 | 0.931 0.955 | 0.953 Freezing 0.748 | 0.627 0.793 | 0.807 0.770 | 0.706 Standing 0.559 | 0.421 0.864 | 0.364 0.679 | 0.390 Macro Average 0.758 | 0.675 0.867 | 0.701 0.801 | 0.683 Table 1. Classification performance metrics with and without eye-tracking data. Each cell shows aggregate results across all cross-validation folds, with eye-tracking (model A, left) and without eye-tracking (model B, right italicized). The incorporation of eye-gaze kinematics improved performance across all metrics and classes, with the largest gain for standing (F1-score: 0.679 vs 0.390). Overall macro-averaged F1-score increased from 0.683 to 0.801. Figure 1. Confusion matrices with and without eye‑gaze data. The multimodal IMU+Gaze model (left) shows improved separation of freezing and standing when compared to the IMU-only model (right). Standing windows misclassified as freezing decreased from 63.6% to 9.1%, and true positive standing classification improved from 36.4% to 86.4%. Values represent the percentage of true labels assigned to each predicted class. Graphical Abstract Wearable systems using inertial sensors frequently misclassify voluntary standing as pathological freezing of gait (right). Adding eye‑gaze kinematics to a deep‑learning model substantially reduces this error (left): standing misclassified as freezing falls from 63.6% to 9.1% in our dataset. The resulting improvement in specificity should aid automated monitoring and objective assessment, and could inform future assistive cueing pending validation in daily‑life settings and causal, low‑latency deployment. Supplementary Material File (table1.docx) Download 15.47 KB Information & Authors Information Version history V1 Version 1 23 August 2025 Peer review timeline Published European Journal of Neuroscience Version of Record 27 Mar 2026 Published Copyright This work is licensed under a Non Exclusive No Reuse License. Collection European Journal of Neuroscience Keywords digital biomarkers eye tracking inertial measurement unit machine learning parkinson's disease Authors Affiliations Christopher Pulliam 0000-0002-4623-6022 [email protected] Case Western Reserve University View all articles by this author Jinxin Chen Case Western Reserve University View all articles by this author James Liao 0000-0001-6605-9855 Cleveland Clinic Neurological Institute View all articles by this author Metrics & Citations Metrics Article Usage 328 views 229 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Christopher Pulliam, Jinxin Chen, James Liao. Integrating Eye Tracking and Inertial Sensing for Enhanced Freezing of Gait Detection in Parkinson's Disease. Authorea . 23 August 2025. DOI: https://doi.org/10.22541/au.175597775.50778140/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.175597775.50778140/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9fed21217d7cdf88',t:'MTc3OTI5ODg1Nw=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();