{"paper_id":"39ba8a2a-9e20-4cd3-81fd-b342d3bb6e06","body_text":"Smartphone-Derived Ocular Motor Biomarkers Enable AI to Assess Neurodegeneration in Huntington’s Disease | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Smartphone-Derived Ocular Motor Biomarkers Enable AI to Assess Neurodegeneration in Huntington’s Disease Leonardo Eleuterio Ariello, Kelvin Wang, David E. Newman-Toker, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8844897/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Digital biomarkers derived from consumer devices offer new opportunities for remote neurological assessment. However, most AI-based approaches depend on large, disease-specific training datasets, limiting their applicability in rare disorders. Large language models (LLMs), trained on broad medical corpora, may enable clinically meaningful inference without disease-specific model training when provided with structured physiological inputs. In this prospective proof-of-concept study, individuals with genetically confirmed Huntington’s disease (HD) and age-matched healthy controls completed an ocular motor assessment using an in-house-developed smartphone application. Quantitative eye movement metrics were validated against expert neurologist ratings and subsequently provided to LLMs using a structured prompt. Models generated an AI-assigned HD probability score (HAIPS) based exclusively on ocular motor data. Twenty-six participants were included. Smartphone-derived metrics showed strong agreement with clinical ratings (Spearman ρ 0.76–0.95; all p < 0.001). HAIPS reliably discriminated individuals with HD from controls (AUC 0.879–0.944), with no significant differences across models. Among HD participants, higher HAIPS correlated with established motor and cognitive measures (Spearman ρ 0.74–0.86; all p < 0.01). These findings demonstrate that LLMs can generate clinically meaningful probabilistic assessments of HD from smartphone-derived ocular motor data without disease-specific training, highlighting a scalable framework for AI-supported assessment in neurodegenerative disorders. Mobile Health Digital Biomarkers Huntington’s disease eye tracking artificial intelligence large language models Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction As the burden of diseases continues to expand, accurate and timely assessment has become an increasing bottleneck to care. 1 This challenge is particularly pronounced in rare disorders such as Huntington’s disease (HD), where clinical evaluation remains largely dependent on in-person examinations performed by fellowship-trained neurologists. 2 The limited availability of these specialists, together with their uneven geographic distribution, creates substantial practical and geographic barriers to optimal care. 3,4 Recent advances in artificial intelligence (AI) and machine learning (ML) have begun to address these challenges, enabling remote assessment using consumer devices such as smartphones. 5–7 However, most existing approaches rely on large, disease-specific datasets, and supervised training pipelines. 8 These requirements limit scalability in rare diseases such as HD. 9 Moreover, disease-specific model training and fine-tuning are resource-intensive and vulnerable to rapid obsolescence as model architectures evolve. 10 General-purpose large language models (LLMs) offer a potential alternative to this paradigm. 11 Through large-scale pretraining, LLMs encode extensive latent medical and physiological knowledge. 12 We hypothesized that, when conceptualized as inference engines rather than trained classifiers, supplying LLMs with a compact set of physiologically grounded biomarkers would enable clinically meaningful probabilistic judgments even in rare diseases. Eye movements represent an especially well-suited physiological signal for this purpose. As in other neurodegenerative diseases, in HD, ocular motor abnormalities are well characterized, emerging early in the disease course and remaining informative across all stages of progression. 13 Moreover, eye movement assessment is rapid to perform, requires minimal instruction, and is tightly linked to specific neural circuits, making it particularly well suited for objective digital phenotyping. 14 In this proof-of-concept study, we developed a custom smartphone application that captures in real time ocular motor features and tested whether LLMs can leverage ocular motor information to generate clinically meaningful, probabilistic inference under a strictly zero-shot framework. Using HD as a model condition, this work explores a scalable, physiology-driven approach to objective disease assessment across diverse neurological conditions. Methods Study Design and Participants We conducted a prospective case–control study including individuals with genetically confirmed HD and matched healthy controls. HD participants were recruited (May-October 2025) from the Johns Hopkins Huntington’s Disease Multidisciplinary Clinic and underwent standardized clinical assessment on the same day, including the Unified Huntington’s Disease Rating Scale (UHDRS) motor score, 15 Montreal Cognitive Assessment (MoCA) and Total Functional Capacity (TFC). Control participants were recruited on a voluntary basis from hospital staff and the local community. All controls underwent a screening process, including medical history, to exclude neurological, psychiatric, or ophthalmologic conditions. Controls were age-matched at the group level to participants with HD. Smartphone Ocular Motor Assessment Participants completed a 10-minute ocular motor assessment using a custom, in-house-developed iOS application running on an iPhone 14 (Apple Inc., Cupertino, CA). The device was positioned approximately 35 cm from the participant’s face and could be held by the participant, a caregiver, or the examiner depending on comfort and motor ability. The ocular motor battery included five ocular motor tasks used in routine clinical evaluation (Fig. 1 ). Signal Processing and Feature Extraction Smartphones captured raw gaze-position traces that were exported in CSV format and analyzed with a standard pipeline ( Supplementary Material 1) . Two main forms of analysis were extracted from the five ocular motor domains: (1) Trajectory-based metrics: Applied to visually guided saccades and smooth pursuit, each recorded eye-movement trajectory was amplitude-normalized and aligned to a task-specific physiological template (pulse–step function for saccades; sinusoidal function for smooth pursuit). A template-deviation score was then computed based on the area under the deviation curve (AUC), quantifying the cumulative deviation of the recorded trajectory from the expected physiological pattern. Importantly, higher AUC values indicate greater deviation from normal eye-movement physiology and therefore worse ocular motor performance, whereas lower AUC values reflect trajectories more closely matching normal physiological behavior. (2) Event-based metrics: Derived directly from discrete behaviors: (i) anti-saccade error rate; (ii) Self-paced saccade count; (iii) fixation deviation rate (> 2°). Neurologist–Smartphone Convergent Validity Automated smartphone-derived eye movement metrics were compared with clinician-rated ocular motor assessments performed by a masked fellowship-trained eye movement neurologist (D.P.W.R). Clinical ratings were based on standard UHDRS eye movement subscores, 15 assigned through review of video recordings obtained during in-person clinical visits. Convergent validity between smartphone metrics and ordinal clinician-assigned severity scores was assessed using Spearman rank correlation coefficients. LLM Prompting and Zero-Shot probability Estimation Automated prompts were generated programmatically using a standardized template ( Supplementary Material 2 ) and participant-level eye movement metrics extracted from the smartphone assessment. Prompts were submitted to four general-purpose LLMs (GPT-5.1, DeepSeek-R1, Gemini, and Claude) with no model fine-tuning, task-specific training examples, or diagnostic labels provided within the prompt. For each participant, models were instructed to (i) interpret the numerical ocular motor metrics, (ii) output AI-Assigned HD Probability Score (HAIPS) on a 0–100 scale, and (iii) assign a categorical classification (“Likely HD,” “Uncertain,” or “Likely Control”) based solely on the provided inputs ( Supplementary Material 2 ). Model inference parameters were set to vendor-default values for temperature and other decoding settings (e.g., top-p and maximum output length), and were held constant across participants within each. 16 Per-participant outputs for all models are reported in Supplementary Table 1 ). A flowchart summarizing the experimental pipeline is shown in Fig. 2 . Statistical analysis All statistical analyses were two-sided, with a significance threshold of α = 0.05. Data distributions were assessed by visual inspection and the Shapiro–Wilk test. Given the small sample size and non-normal distributions, non-parametric methods were applied throughout. HAIPS was analyzed both as continuous variables and as categorical classifications. Model discrimination performance was evaluated using receiver operating characteristic (ROC) analysis, with differences in AUC compared using the DeLong method. Associations between HAIPS and clinical measures of disease severity were assessed using Spearman rank correlation coefficients. All analyses were performed in SPSS 25 and Python (version 3.11). Results Participant Characteristics Twenty-six individuals were enrolled, including 13 participants with genetically confirmed HD and 13 age- and sex-matched healthy controls. All participants completed the smartphone-based assessment without difficulty. The HD and control groups were comparable in sex distribution (HD: 6 males/7 females; controls: 8 males/5 females) and age. Median age was 48 years (IQR 12) in the HD group and 53 years (IQR 27) in controls. Among participants with HD, median CAG repeat length was 44 (IQR 4). Median UHDRS motor score was 35 (IQR 33), median TFC score was 7 (IQR 8), and median MoCA score was 23 (IQR 5). Agreement of Smartphone Metrics with Neurologist Scoring Smartphone metrics showed strong agreement with ocular motor assessments performed by an expert neurologist (Fig. 3 ). Spearman rank correlation analyses demonstrated robust positive associations across all evaluated domains, including vertical saccades (ρ = 0.93), horizontal saccades (ρ = 0.76), vertical smooth pursuit (ρ = 0.94), and horizontal smooth pursuit (ρ = 0.95), with all correlations reaching statistical significance (p < 0.001). The number of fixation losses and anti-saccade errors identified by the smartphone application showed exact correspondence with physician-based observations. Across both saccadic and smooth pursuit measures, smartphone metrics increased progressively with higher clinical severity scores. Median values differed systematically across severity categories, with minimal overlap between the lowest (score 0) and highest (score 4) groups. Detailed median values and interquartile ranges for each ocular motor domain across severity levels are provided in Supplementary Table 2. LLMs Diagnosis HD with Robust Discrimination Smartphone-derived ocular motor metrics were provided to four general-purpose LLMs using a standardized prompt ( Supplementary Material 2 ). All models tested were able to distinguish HD from age-matched controls based only on ocular motor analysis. Participants with HD consistently received higher HAIPS, with minimal overlap between probability distributions Fig. 4 . Gemini and GPT-5.1 showed the highest performance (AUC 0.944, 95% CI 0.83–1.00 and 0.935, 95% CI 0.82–1.00, respectively), while Claude (AUC 0.885, 95% CI 0.71–1.00) and DeepSeek R1 (AUC 0.879, 95% CI 0.71–0.98) performed only slightly lower. DeLong comparisons between the highest- and lowest-performing models yielded p > 0.05 for all pairwise tests, indicating that the four LLMs converged on a broadly similar physiological decision boundary, and this approach is model agnostic. Association of HAIPS with Clinical Measures of Disease Severity The HAIPS, derived exclusively from LLM-based analysis of smartphone eye movement data, was compared with standard clinical measures of cognition, functional capacity, and motor impairment obtained during in-person examination. Among participants, higher HAIPS values were strongly associated with poorer cognitive performance on the MoCA (ρ = − 0.86, p < 0.001), reduced functional capacity on the TFC scale (ρ = − 0.74, p = 0.003), and greater motor impairment on the UHDRS motor scale (ρ = 0.85, p < 0.001) (Fig. 5 ). Demonstrating that HAIPS captures not only diagnostic status from the ocular motor data but also varies proportionally with established measures of disease severity across clinical domains. Discussion In this proof-of-concept study, we show that general-purpose LLMs can extract clinically meaningful information from ocular motor data acquired using a conventional smartphone application. Using eye movements features alone, the models generated a probability score, that reliably distinguished HD from healthy controls and scaled with disease severity. Our findings indicate that the models captured not only the presence of ocular motor abnormalities, but also their graded clinical relevance. Initially we demonstrated clinical validity by showing strong concordance between smartphone ocular motor metrics and clinical assessments performed by trained eye movement specialists. The application generated physiologically plausible eye movement traces and summary metrics that aligned with well established clinical measures in HD (UHDRS motor score). Translating smartphone recordings into scalable and interpretable clinical inference has traditionally required expert review, curated feature extraction, or supervised ML pipelines. 17 These approaches are labor-intensive and depend on large, well-annotated datasets, limiting their applicability in specific domains and in rare diseases such as HD, where data scarcity is inherent. 18 To address this bottleneck we leveraged LLMs as general-purpose inference engines. 11 First, unlike conventional classifiers, LLMs do not require disease-specific training, feature reweighting, or retraining as model architectures evolve. Second, LLMs can operate directly in a zero-shot settings (e.g., without receiving any prior training examples or specific, labeled data), mapping structured, physiologically grounded eye-movement metrics to probabilistic clinical assessments without supervised optimization or model fine-tuning. The zero-shot nature of our approach is particularly important for rare neurological diseases, where large disease-specific training datasets are often unavailable. 9,19 By operating directly on a limited set of objective, physiologically meaningful inputs, LLMs generated clinically relevant probability estimates without retraining. Notably, all LLMs tested showed similar performance, demonstrating model-agnostic behavior and indicating that diagnostic inference was driven primarily by the structure of the ocular motor data rather than by idiosyncratic properties of any single model. 20 Consistent with this model-agnostic behavior, all models achieved strong discrimination between HD and control participants, with AUC values exceeding 0.9. Moreover, the resulting HAIPS closely tracked clinical impairment across cognitive and motor domains (MoCA, UHDRS motor score and TFC), indicating that the models captured not only the presence of ocular motor abnormalities, but also their severity. While discrimination metrics may be influenced by the limited sample size and the comparison with healthy controls, these findings suggest that ocular motor inputs allow LLMs to unlock latent representations of normal and abnormal physiology, enabling clinically meaningful probabilistic inference via prompt-based reasoning alone. 21 These results are particularly timely in an era of rapid expansion of smartphone neurological assessment. Personal devices have been used to capture motor, gait, speech, and cognitive features across multiple neurodegenerative diseases. 22,23 However, current approaches have pitfalls, by either requiring multiple body sensors, relying on task-specific models, or depending on effortful tasks influenced by educational level or task comprehension. 24,25 Those facts limit scalability and increase variability in real-world settings. 7 In contrast, eye movements offer distinct advantages: they are rapid and easy to assess, relatively independent of patient effort and educational level, and less influenced by peripheral biomechanics. Moreover, ocular motor control is grounded in well-characterized neural circuits, providing a robust physiological substrate for neurological assessment. 26 We showed that when captured using consumer smartphones and interpreted through a LLM-based inference system, ocular motor signals can be transformed into clinically meaningful, probabilistic representations of neurological dysfunction (and, potentially, neurological disease diagnosis) without reliance on disease-specific training. While this study focused on HD, its implications extend beyond a single disorder. Eye movements abnormalities are prevalent across a wide range of neurodegenerative conditions, including dementias, parkinsonian syndromes, motor neuron diseases, and cerebellar disorders. 26 The ability to derive disease-relevant signals from eye movements using AI approach points to a broadly applicable strategy for objective digital phenotyping across neurological diseases. Integrating ocular motor metrics with other described smartphone-derived modalities, such as motor performance, speech, and cognitive measures, may further improve sensitivity to disease burden, progression, and therapeutic response. 27 As disease-modifying therapies emerge for HD and other neurodegenerative disorders, scalable and objective tools for remote assessment and longitudinal monitoring are increasingly important. 28,29 Such digital biomarkers may ultimately serve for early detection or treatment monitoring. Several limitations should be considered when interpreting these findings. First, the sample size was modest and derived from a single tertiary center, consistent with the exploratory nature of this study. Larger, demographically diverse cohorts will be necessary to refine estimates of diagnostic accuracy, quantify population-level variability. Second, real-world deployment will require validation across device generations, user-handling variability, as well as testing during unsupervised home use. 30 Third, the cross-sectional design precludes conclusions regarding sensitivity to longitudinal change, premanifest detection, or treatment effects, domains that will be critical for translation into clinical trials and disease-monitoring contexts. Finally, it remains unknown how well this approach might perform with more challenging diagnostic tasks (e.g., HD vs. Alzheimer’s disease, as opposed to healthy controls). In conclusion, eye movement recordings acquired using a standard consumer smartphone contain physiologically informative signals that can support clinically meaningful inference. When interpreted by general-purpose LLMs under a strictly zero-shot framework, these signals yield probabilistic disease estimates that reflect both disease presence and severity. By pairing a disease-agnostic inference model with a physiological marker that reflects distributed neural circuit function, this work provides proof-of-concept for a scalable framework with broad potential across neurological diseases and other physiology-driven clinical domains. Data availability The datasets generated and analyzed during the current study are not publicly available due to privacy and ethical restrictions related to the sharing of identifiable or potentially re-identifiable biometric data, particularly in the context of a rare genetic disorder, but are available from the corresponding author on reasonable request. Declarations Competing Interests David E. Newman-Toker holds an approved patent concerning the use of the EyePhone for tracking eye and head position and has a provisional patent application for the use of the EyePhone in recording saccades and smooth pursuit. The other authors declare no competing interests. Supplementary Material Supplementary material is available at Brain online Ethics approval The study was approved by the Johns Hopkins Medicine Institutional Review Board (IRB00258938). All participants provided written informed consent prior to participation, and all procedures were conducted in accordance with the Declaration of Helsinki. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors Author Contribution L.A. conceived the study, designed the experimental protocol, performed the literature review, conducted data analysis, and wrote the initial draft of the manuscript. K.W. developed the smartphone application, implemented the data-processing pipeline, and contributed to the design and optimization of the ocular motor battery. D.N.T. contributed to model design and provided critical input on neurological interpretation J.B. supervised clinical aspects of the study, oversaw patient recruitment in the Huntington’s disease clinic, and provided critical input on neurological interpretation. D.P.W.R. supervised the AI and computational framework, contributed to model design and prompt engineering, and guided analytical validation. All authors contributed to the interpretation of the data, critically revised the manuscript for important intellectual content, and approved the final version of the manuscript. Acknowledgement The authors thank the patients and their families for their participation in this study. We also acknowledge the support of the clinical and research teams involved in the evaluation and care of individuals with Huntington’s disease. This research did not received specific grant from any funding agency in the public, commercial, or not-for-profit sectors Data Availability The datasets generated and analyzed during the current study are not publicly available due to privacy and ethical restrictions related to the sharing of identifiable or potentially re-identifiable biometric data, particularly in the context of a rare genetic disorder, but are available from the corresponding author on reasonable request. References Huang Y, Li Y, Pan H, Han L. Global, regional, and national burden of neurological disorders in 204 countries and territories worldwide. J Glob Health . 2023;13:04160. doi:10.7189/jogh.13.04160 Bates GP, Dorsey R, Gusella JF, et al. Huntington disease. Nat Rev Dis Primer . 2015;1(1):15005. doi:10.1038/nrdp.2015.5 Dorsey ER, Topol EJ. State of Telehealth. Campion EW, ed. N Engl J Med . 2016;375(2):154-161. doi:10.1056/NEJMra1601705 Bokinni Y. Huntington’s disease: new gene therapy explained. BMJ . 2025;390:r2029. doi:10.1136/bmj.r2029 Dagum P. Digital biomarkers of cognitive function. Npj Digit Med . 2018;1(1):10. doi:10.1038/s41746-018-0018-4 Lim WS, Fan SP, Chiu SI, et al. Smartphone-derived multidomain features including voice, finger-tapping movement and gait aid early identification of Parkinson’s disease. Npj Park Dis . 2025;11(1):111. doi:10.1038/s41531-025-00953-w Yang K, Xiong WX, Liu FT, et al. Objective and quantitative assessment of motor function in Parkinson’s disease—from the perspective of practical applications. Ann Transl Med . 2016;4(5):90-90. doi:10.21037/atm.2016.03.09 Beam AL, Kohane IS. Big Data and Machine Learning in Health Care. JAMA . 2018;319(13):1317. doi:10.1001/jama.2017.18391 Banerjee J, Taroni JN, Allaway RJ, Prasad DV, Guinney J, Greene C. Machine learning in rare disease. Nat Methods . 2023;20(6):803-814. doi:10.1038/s41592-023-01886-z Sculley D, Holt G, Golovin D, et al. Hidden Technical Debt in Machine Learning Systems. NIPS . Published online January 2015:2494-2502. Bommasani R, Hudson DA, Adeli E, et al. On the Opportunities and Risks of Foundation Models. arXiv . Preprint posted online 2021. doi:10.48550/ARXIV.2108.07258 Singhal K, Azizi S, Tu T, et al. Large language models encode clinical knowledge. Nature . 2023;620(7972):172-180. doi:10.1038/s41586-023-06291-2 Leigh RJ, Zee DS. The Neurology of Eye Movements . 5th ed. Oxford university press; 2015. Eye Tracking: A Comprehensive Guide to Methods and Measures . Oxford university press; 2011. Unified Huntington’s disease rating scale: Reliability and consistency. Mov Disord . 1996;11(2):136-142. doi:10.1002/mds.870110204 Wei J, Bosma M, Zhao VY, et al. Finetuned Language Models Are Zero-Shot Learners. arXiv . Preprint posted online 2021. doi:10.48550/ARXIV.2109.01652 Ellis RJ, Sander RM, Limon A. Twelve key challenges in medical machine learning and solutions. Intell-Based Med . 2022;6:100068. doi:10.1016/j.ibmed.2022.100068 Ngiam KY, Khor IW. Big data and machine learning algorithms for health-care delivery. Lancet Oncol . 2019;20(5):e262-e273. doi:10.1016/S1470-2045(19)30149-4 Decherchi S, Pedrini E, Mordenti M, Cavalli A, Sangiorgi L. Opportunities and Challenges for Machine Learning in Rare Diseases. Front Med . 2021;8:747612. doi:10.3389/fmed.2021.747612 Liang P, Bommasani R, Lee T, et al. Holistic Evaluation of Language Models. Published online 2022. doi:10.48550/ARXIV.2211.09110 Lake BM, Ullman TD, Tenenbaum JB, Gershman SJ. Building machines that learn and think like people. Behav Brain Sci . 2017;40:e253. doi:10.1017/S0140525X16001837 Bot BM, Suver C, Neto EC, et al. The mPower study, Parkinson disease mobile data collected using ResearchKit. Sci Data . 2016;3(1):160011. doi:10.1038/sdata.2016.11 Koerner J, Zou E, Karl JA, et al. Towards scalable screening for the early detection of Parkinson’s disease: validation of an iPad-based eye movement assessment system against a clinical-grade eye tracker. Npj Park Dis . 2025;11(1):233. doi:10.1038/s41531-025-01079-9 Meng X, D’Arcy C. Education and Dementia in the Context of the Cognitive Reserve Hypothesis: A Systematic Review with Meta-Analyses and Qualitative Analyses. Laks J, ed. PLoS ONE . 2012;7(6):e38268. doi:10.1371/journal.pone.0038268 Del Din S, Elshehabi M, Galna B, et al. Gait analysis with wearables predicts conversion to Parkinson disease. Ann Neurol . 2019;86(3):357-367. doi:10.1002/ana.25548 Antoniades CA, Kennard C. Ocular motor abnormalities in neurodegenerative disorders. Eye . 2015;29(2):200-207. doi:10.1038/eye.2014.276 Sun Y meng, Wang Z yun, Liang Y yuan, Hao C wei, Shi C he. Digital biomarkers for precision diagnosis and monitoring in Parkinson’s disease. Npj Digit Med . 2024;7(1):218. doi:10.1038/s41746-024-01217-2 eClinicalMedicine. A new hope for patients with Huntington’s disease? eClinicalMedicine . 2025;88:103612. doi:10.1016/j.eclinm.2025.103612 Cummings J, Lee G, Nahed P, et al. Alzheimer’s disease drug development pipeline: 2022. Alzheimers Dement Transl Res Clin Interv . 2022;8(1):e12295. doi:10.1002/trc2.12295 Goetz L, Seedat N, Vandersluis R, Van Der Schaar M. Generalization—a key challenge for responsible AI in patient-facing clinical applications. Npj Digit Med . 2024;7(1):126. doi:10.1038/s41746-024-01127-3 Additional Declarations Competing interest reported. David E. Newman-Toker holds an approved patent concerning the use of the EyePhone for tracking eye and head position and has a provisional patent application for the use of the EyePhone in recording saccades and smooth pursuit. The other authors declare no competing interests. Supplementary Files SupplementaryMaterial.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {\"props\":{\"pageProps\":{\"initialData\":{\"identity\":\"rs-8844897\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":true,\"archivedVersions\":[],\"articleType\":\"Research Article\",\"associatedPublications\":[],\"authors\":[{\"id\":598914344,\"identity\":\"ad917739-29e9-46f0-b800-47ae143a0557\",\"order_by\":0,\"name\":\"Leonardo Eleuterio Ariello\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAuElEQVRIie3RsQrCMBCA4SuFukSzRvAhIoIIFnwVp3Z1Ko6ZnPoAeZyErGmfQZfOGTsoeHZx7GUTzA+BLB+5cACp1A+2wmPgcmAcPF4LAikmIsVmrWIIICmlYVSy6B4mSMF2tjMQGkcgrJZWI9mb/pzpnkKgAscm4mW+vFEIH8A9P4MpJC8SEfgKfp9JQJLRyAC2RSJwMNv29TzhvMrDeC1PXPvtfWyO8+SbMLjTuLiKBKlUKvU3vQHlITY3Gb7r4wAAAABJRU5ErkJggg==\",\"orcid\":\"\",\"institution\":\"Johns Hopkins University\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"Leonardo\",\"middleName\":\"Eleuterio\",\"lastName\":\"Ariello\",\"suffix\":\"\"},{\"id\":598914345,\"identity\":\"404124ed-826e-42f2-a160-711bf5016c0e\",\"order_by\":1,\"name\":\"Kelvin Wang\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Johns Hopkins University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Kelvin\",\"middleName\":\"\",\"lastName\":\"Wang\",\"suffix\":\"\"},{\"id\":598914346,\"identity\":\"d7eae2c3-afbd-48c6-8f4a-3ea56ba9f2bd\",\"order_by\":2,\"name\":\"David E. Newman-Toker\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Johns Hopkins University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"David\",\"middleName\":\"E.\",\"lastName\":\"Newman-Toker\",\"suffix\":\"\"},{\"id\":598914347,\"identity\":\"ac70a4c8-f53b-40dd-b437-cc658be49658\",\"order_by\":3,\"name\":\"Jee Bang\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Johns Hopkins University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Jee\",\"middleName\":\"\",\"lastName\":\"Bang\",\"suffix\":\"\"},{\"id\":598914348,\"identity\":\"45a56fad-b817-4dbe-9f5f-78a285e5cd5f\",\"order_by\":4,\"name\":\"David P.W. Rastall\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Johns Hopkins University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"David\",\"middleName\":\"P.W.\",\"lastName\":\"Rastall\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2026-02-10 19:53:24\",\"currentVersionCode\":1,\"declarations\":\"\",\"doi\":\"10.21203/rs.3.rs-8844897/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-8844897/v1\",\"draftVersion\":[],\"editorialEvents\":[],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":104168959,\"identity\":\"ee3c0b7f-df46-46a8-9fc4-538fc341c081\",\"added_by\":\"auto\",\"created_at\":\"2026-03-08 14:37:02\",\"extension\":\"png\",\"order_by\":1,\"title\":\"Figure 1\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":692202,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eSmartphone-based ocular motor paradigms and extraction of physiologically grounded quantitative metrics. \\u003c/strong\\u003eRepresentative eye- movement recordings obtained using a custom smartphone application from a neurologically healthy control and an individual with Huntington’s disease across the five ocular motor paradigms included in the assessment battery: visually guided saccades, self-paced saccades, smooth pursuit, anti-saccades, and fixation with distractors. Raw gaze-position traces were recorded using the smartphone infrared camera and processed through a standardized signal-processing pipeline (see Methods and Supplementary Material 1). For trajectory-based tasks (visually guided saccades and smooth pursuit), recorded eye-movement trajectories were amplitude-normalized and aligned to task-specific physiological templates (pulse–step function for saccades; sinusoidal function for smooth pursuit). Quantitative deviation from expected physiology was summarized using a template-deviation score based on the area under the deviation curve (AUC), where higher values indicate greater deviation and worse ocular motor performance. For event-based tasks, discrete behavioral metrics were extracted directly, including anti-saccade error rate, self-paced saccade count, and fixation deviation rate.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure01.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8844897/v1/1644854438fc95a6bbda973f.png\"},{\"id\":104168960,\"identity\":\"a42c7164-207a-4aa9-8b62-861cb30fcce6\",\"added_by\":\"auto\",\"created_at\":\"2026-03-08 14:37:02\",\"extension\":\"png\",\"order_by\":2,\"title\":\"Figure 2\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":635403,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eSmartphone ocular motor assessment and AI interpretation workflow. \\u003c/strong\\u003eIndividuals with Huntington’s disease and age-matched controls completed a brief smartphone \\u0026nbsp;eye-movement assessment (10 minutes). Quantitative features extracted from saccade, smooth pursuit, fixation, and other ocular motor tasks were automatically formatted into a structured prompt and processed by a large language model (LLM), which produced an AI-Assigned HD Probability Score (HAIPS), based solely on the ocular motor data.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure02.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8844897/v1/9b688c4aa3249ce62119f661.png\"},{\"id\":104168964,\"identity\":\"f2d68e2f-fd3a-4dd3-9067-e644a052257a\",\"added_by\":\"auto\",\"created_at\":\"2026-03-08 14:37:02\",\"extension\":\"png\",\"order_by\":3,\"title\":\"Figure 3\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":551104,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eConvergent validity between clinician-assigned ocular motor severity and smartphone-derived quantitative metrics.\\u003c/strong\\u003e Violin plots show the distribution of continuous smartphone-derived ocular motor impairment metrics across increasing clinician-assigned severity scores (0–4) for four domains: vertical saccades, horizontal saccades, vertical smooth pursuit, and horizontal smooth pursuit, among participants with HD (n=13). Severity scores were independently assigned by an experienced neurologist masked to all quantitative outputs, based solely on visual inspection of the smartphone eye-movement recordings, using the ocular motor subitems of the Unified Huntington’s Disease Rating Scale (UHDRS). For smooth pursuit, scores were defined as:\\u003cstrong\\u003e \\u003c/strong\\u003e0 = complete (normal) pursuit; 1 = jerky pursuit; 2 = interrupted pursuit with full range; 3 = incomplete pursuit range; 4 = inability to pursue the target. For\\u003cstrong\\u003e \\u003c/strong\\u003esaccades\\u003cstrong\\u003e, \\u003c/strong\\u003eseverity scores reflected an integrated clinical assessment of\\u003cstrong\\u003e \\u003c/strong\\u003esaccade initiation and saccade velocity, defined as:\\u003cstrong\\u003e \\u003c/strong\\u003e0 = normal initiation and velocity; 1 = increased initiation latency or mild saccadic slowing; 2 = suppressible blinks or head movements required to initiate saccades or moderate slowing; 3 = insuppressible head movements or severely slowed saccades with preserved range; 4 = inability to initiate saccades or incomplete saccadic range\\u003cstrong\\u003e. \\u003c/strong\\u003eSmartphone-derived metrics reflect normalized deviation from expected physiological eye-movement templates, with higher values indicating greater ocular motor impairment. Horizontal bars denote median values and interquartile ranges (IQRs).\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure03.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8844897/v1/4261d1778475d31598b90397.png\"},{\"id\":104404448,\"identity\":\"4c2d39cd-ab4f-49b3-88a7-6b2350de01ee\",\"added_by\":\"auto\",\"created_at\":\"2026-03-11 12:20:18\",\"extension\":\"png\",\"order_by\":4,\"title\":\"Figure 4\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":674187,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eROC Curves Demonstrating AI-Based Discrimination of Huntington’s Disease: \\u003c/strong\\u003eReceiver operating characteristic (ROC) curves illustrating the ability of four LLMs to distinguish individuals with Huntington’s disease (HD) from age-matched controls using only smartphone ocular motor metrics. Each panel displays the ROC curve for a single model along with its corresponding area under the curve (AUC).\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure04.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8844897/v1/2a96e877b689fb329c06a9b5.png\"},{\"id\":104403759,\"identity\":\"b2c66404-29ba-4e01-a284-63d6753d00da\",\"added_by\":\"auto\",\"created_at\":\"2026-03-11 12:18:59\",\"extension\":\"png\",\"order_by\":5,\"title\":\"Figure 5\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":306544,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eAssociations Between Clinical Severity and AI-Assigned HD Probability Score (HAIPS): \\u003c/strong\\u003eScatterplots illustrating the relationship between HAIPS and clinical severity across three domains. \\u003cstrong\\u003e(Left)\\u003c/strong\\u003eLower cognitive performance on the Montreal Cognitive Assessment (MoCA) was associated with higher HD probability. \\u003cstrong\\u003e(Middle)\\u003c/strong\\u003e Reduced functional capacity, measured by the Total Functional Capacity scale (TFC), corresponded to higher HAIPS. \\u003cstrong\\u003e(Right)\\u003c/strong\\u003e Greater motor impairment on the UHDRS-Motor Score was associated with higher HAIPS. Each marker represents one participant with HD. Shaded regions denote 95% confidence intervals around the fitted monotonic trend. Across all domains, worse cognitive, functional, and motor performance aligned with increased AI-assigned probability of HD (HAIPS), indicating that the model captures continuous variation in disease burden.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure05.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8844897/v1/564d22acb036fe5560d6223e.png\"},{\"id\":105566718,\"identity\":\"a8533a06-d96b-4a5e-8836-04b6e895482c\",\"added_by\":\"auto\",\"created_at\":\"2026-03-27 12:57:06\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":3292181,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"manuscript.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8844897/v1/74790ec8-82b5-457e-8495-edd1101279ed.pdf\"},{\"id\":104168963,\"identity\":\"60f69f31-27d4-49ac-b3fd-e90239986dff\",\"added_by\":\"auto\",\"created_at\":\"2026-03-08 14:37:02\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":234894,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryMaterial.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8844897/v1/070984d59c5f5925d1b62d5e.pdf\"}],\"financialInterests\":\"Competing interest reported. David E. Newman-Toker holds an approved patent concerning the use of the EyePhone for tracking eye and head position and has a provisional patent application for the use of the EyePhone in recording saccades and smooth pursuit. The other authors declare no competing interests.\",\"formattedTitle\":\"Smartphone-Derived Ocular Motor Biomarkers Enable AI to Assess Neurodegeneration in Huntington’s Disease\",\"fulltext\":[{\"header\":\"Introduction\",\"content\":\"\\u003cp\\u003eAs the burden of diseases continues to expand, accurate and timely assessment has become an increasing bottleneck to care.\\u003csup\\u003e1\\u003c/sup\\u003e This challenge is particularly pronounced in rare disorders such as Huntington\\u0026rsquo;s disease (HD), where clinical evaluation remains largely dependent on in-person examinations performed by fellowship-trained neurologists.\\u003csup\\u003e2\\u003c/sup\\u003e The limited availability of these specialists, together with their uneven geographic distribution, creates substantial practical and geographic barriers to optimal care.\\u003csup\\u003e3,4\\u003c/sup\\u003e\\u003c/p\\u003e \\u003cp\\u003eRecent advances in artificial intelligence (AI) and machine learning (ML) have begun to address these challenges, enabling remote assessment using consumer devices such as smartphones.\\u003csup\\u003e5\\u0026ndash;7\\u003c/sup\\u003e However, most existing approaches rely on large, disease-specific datasets, and supervised training pipelines.\\u003csup\\u003e8\\u003c/sup\\u003e These requirements limit scalability in rare diseases such as HD.\\u003csup\\u003e9\\u003c/sup\\u003e Moreover, disease-specific model training and fine-tuning are resource-intensive and vulnerable to rapid obsolescence as model architectures evolve.\\u003csup\\u003e10\\u003c/sup\\u003e\\u003c/p\\u003e \\u003cp\\u003eGeneral-purpose large language models (LLMs) offer a potential alternative to this paradigm.\\u003csup\\u003e11\\u003c/sup\\u003e Through large-scale pretraining, LLMs encode extensive latent medical and physiological knowledge.\\u003csup\\u003e12\\u003c/sup\\u003e We hypothesized that, when conceptualized as inference engines rather than trained classifiers, supplying LLMs with a compact set of physiologically grounded biomarkers would enable clinically meaningful probabilistic judgments even in rare diseases.\\u003c/p\\u003e \\u003cp\\u003eEye movements represent an especially well-suited physiological signal for this purpose. As in other neurodegenerative diseases, in HD, ocular motor abnormalities are well characterized, emerging early in the disease course and remaining informative across all stages of progression.\\u003csup\\u003e13\\u003c/sup\\u003e Moreover, eye movement assessment is rapid to perform, requires minimal instruction, and is tightly linked to specific neural circuits, making it particularly well suited for objective digital phenotyping.\\u003csup\\u003e14\\u003c/sup\\u003e\\u003c/p\\u003e \\u003cp\\u003eIn this proof-of-concept study, we developed a custom smartphone application that captures in real time ocular motor features and tested whether LLMs can leverage ocular motor information to generate clinically meaningful, probabilistic inference under a strictly zero-shot framework. Using HD as a model condition, this work explores a scalable, physiology-driven approach to objective disease assessment across diverse neurological conditions.\\u003c/p\\u003e\"},{\"header\":\"Methods\",\"content\":\"\\u003cdiv id=\\\"Sec3\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eStudy Design and Participants\\u003c/h2\\u003e \\u003cp\\u003eWe conducted a prospective case\\u0026ndash;control study including individuals with genetically confirmed HD and matched healthy controls. HD participants were recruited (May-October 2025) from the Johns Hopkins Huntington\\u0026rsquo;s Disease Multidisciplinary Clinic and underwent standardized clinical assessment on the same day, including the Unified Huntington\\u0026rsquo;s Disease Rating Scale (UHDRS) motor score,\\u003csup\\u003e15\\u003c/sup\\u003e Montreal Cognitive Assessment (MoCA) and Total Functional Capacity (TFC). Control participants were recruited on a voluntary basis from hospital staff and the local community. All controls underwent a screening process, including medical history, to exclude neurological, psychiatric, or ophthalmologic conditions. Controls were age-matched at the group level to participants with HD.\\u003c/p\\u003e \\u003c/div\\u003e\\n\\u003ch3\\u003eSmartphone Ocular Motor Assessment\\u003c/h3\\u003e\\n\\u003cp\\u003eParticipants completed a 10-minute ocular motor assessment using a custom, in-house-developed iOS application running on an iPhone 14 (Apple Inc., Cupertino, CA). The device was positioned approximately 35 cm from the participant\\u0026rsquo;s face and could be held by the participant, a caregiver, or the examiner depending on comfort and motor ability. The ocular motor battery included five ocular motor tasks used in routine clinical evaluation (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e).\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e\\n\\u003ch3\\u003eSignal Processing and Feature Extraction\\u003c/h3\\u003e\\n\\u003cp\\u003eSmartphones captured raw gaze-position traces that were exported in \\u003cem\\u003eCSV\\u003c/em\\u003e format and analyzed with a standard pipeline (\\u003cem\\u003eSupplementary Material 1)\\u003c/em\\u003e. Two main forms of analysis were extracted from the five ocular motor domains:\\u003c/p\\u003e \\u003cp\\u003e(1) Trajectory-based metrics: Applied to visually guided saccades and smooth pursuit, each recorded eye-movement trajectory was amplitude-normalized and aligned to a task-specific physiological template (pulse\\u0026ndash;step function for saccades; sinusoidal function for smooth pursuit). A template-deviation score was then computed based on the area under the deviation curve (AUC), quantifying the cumulative deviation of the recorded trajectory from the expected physiological pattern. Importantly, higher AUC values indicate greater deviation from normal eye-movement physiology and therefore worse ocular motor performance, whereas lower AUC values reflect trajectories more closely matching normal physiological behavior.\\u003c/p\\u003e \\u003cp\\u003e(2) Event-based metrics: Derived directly from discrete behaviors: (i) anti-saccade error rate; (ii) Self-paced saccade count; (iii) fixation deviation rate (\\u0026gt;\\u0026thinsp;2\\u0026deg;).\\u003c/p\\u003e\\n\\u003ch3\\u003eNeurologist–Smartphone Convergent Validity\\u003c/h3\\u003e\\n\\u003cp\\u003eAutomated smartphone-derived eye movement metrics were compared with clinician-rated ocular motor assessments performed by a masked fellowship-trained eye movement neurologist (D.P.W.R). Clinical ratings were based on standard UHDRS eye movement subscores,\\u003csup\\u003e15\\u003c/sup\\u003e assigned through review of video recordings obtained during in-person clinical visits. Convergent validity between smartphone metrics and ordinal clinician-assigned severity scores was assessed using Spearman rank correlation coefficients.\\u003c/p\\u003e\\n\\u003ch3\\u003eLLM Prompting and Zero-Shot probability Estimation\\u003c/h3\\u003e\\n\\u003cp\\u003eAutomated prompts were generated programmatically using a standardized template (\\u003cem\\u003eSupplementary Material 2\\u003c/em\\u003e) and participant-level eye movement metrics extracted from the smartphone assessment. Prompts were submitted to four general-purpose LLMs (GPT-5.1, DeepSeek-R1, Gemini, and Claude) with no model fine-tuning, task-specific training examples, or diagnostic labels provided within the prompt. For each participant, models were instructed to (i) interpret the numerical ocular motor metrics, (ii) output AI-Assigned HD Probability Score (HAIPS) on a 0\\u0026ndash;100 scale, and (iii) assign a categorical classification (\\u0026ldquo;Likely HD,\\u0026rdquo; \\u0026ldquo;Uncertain,\\u0026rdquo; or \\u0026ldquo;Likely Control\\u0026rdquo;) based solely on the provided inputs (\\u003cem\\u003eSupplementary Material 2\\u003c/em\\u003e).\\u003c/p\\u003e \\u003cp\\u003eModel inference parameters were set to vendor-default values for temperature and other decoding settings (e.g., top-p and maximum output length), and were held constant across participants within each.\\u003csup\\u003e16\\u003c/sup\\u003e Per-participant outputs for all models are reported in \\u003cem\\u003eSupplementary Table\\u0026nbsp;1\\u003c/em\\u003e). A flowchart summarizing the experimental pipeline is shown in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cdiv id=\\\"Sec8\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eStatistical analysis\\u003c/h2\\u003e \\u003cp\\u003eAll statistical analyses were two-sided, with a significance threshold of α\\u0026thinsp;=\\u0026thinsp;0.05. Data distributions were assessed by visual inspection and the Shapiro\\u0026ndash;Wilk test. Given the small sample size and non-normal distributions, non-parametric methods were applied throughout. HAIPS was analyzed both as continuous variables and as categorical classifications. Model discrimination performance was evaluated using receiver operating characteristic (ROC) analysis, with differences in AUC compared using the DeLong method. Associations between HAIPS and clinical measures of disease severity were assessed using Spearman rank correlation coefficients. All analyses were performed in SPSS 25 and Python (version 3.11).\\u003c/p\\u003e \\u003c/div\\u003e\"},{\"header\":\"Results\",\"content\":\"\\u003cdiv id=\\\"Sec10\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eParticipant Characteristics\\u003c/h2\\u003e \\u003cp\\u003e Twenty-six individuals were enrolled, including 13 participants with genetically confirmed HD and 13 age- and sex-matched healthy controls. All participants completed the smartphone-based assessment without difficulty. The HD and control groups were comparable in sex distribution (HD: 6 males/7 females; controls: 8 males/5 females) and age. Median age was 48 years (IQR 12) in the HD group and 53 years (IQR 27) in controls. Among participants with HD, median CAG repeat length was 44 (IQR 4). Median UHDRS motor score was 35 (IQR 33), median TFC score was 7 (IQR 8), and median MoCA score was 23 (IQR 5).\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec11\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eAgreement of Smartphone Metrics with Neurologist Scoring\\u003c/h2\\u003e \\u003cp\\u003eSmartphone metrics showed strong agreement with ocular motor assessments performed by an expert neurologist (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003e). Spearman rank correlation analyses demonstrated robust positive associations across all evaluated domains, including vertical saccades (ρ\\u0026thinsp;=\\u0026thinsp;0.93), horizontal saccades (ρ\\u0026thinsp;=\\u0026thinsp;0.76), vertical smooth pursuit (ρ\\u0026thinsp;=\\u0026thinsp;0.94), and horizontal smooth pursuit (ρ\\u0026thinsp;=\\u0026thinsp;0.95), with all correlations reaching statistical significance (p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001). The number of fixation losses and anti-saccade errors identified by the smartphone application showed exact correspondence with physician-based observations.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003eAcross both saccadic and smooth pursuit measures, smartphone metrics increased progressively with higher clinical severity scores. Median values differed systematically across severity categories, with minimal overlap between the lowest (score 0) and highest (score 4) groups. Detailed median values and interquartile ranges for each ocular motor domain across severity levels are provided in \\u003cem\\u003eSupplementary Table\\u0026nbsp;2.\\u003c/em\\u003e\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec12\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eLLMs Diagnosis HD with Robust Discrimination\\u003c/h2\\u003e \\u003cp\\u003eSmartphone-derived ocular motor metrics were provided to four general-purpose LLMs using a standardized prompt (\\u003cem\\u003eSupplementary Material 2\\u003c/em\\u003e). All models tested were able to distinguish HD from age-matched controls based only on ocular motor analysis. Participants with HD consistently received higher HAIPS, with minimal overlap between probability distributions Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003eGemini and GPT-5.1 showed the highest performance (AUC 0.944, 95% CI 0.83\\u0026ndash;1.00 and 0.935, 95% CI 0.82\\u0026ndash;1.00, respectively), while Claude (AUC 0.885, 95% CI 0.71\\u0026ndash;1.00) and DeepSeek R1 (AUC 0.879, 95% CI 0.71\\u0026ndash;0.98) performed only slightly lower. DeLong comparisons between the highest- and lowest-performing models yielded p\\u0026thinsp;\\u0026gt;\\u0026thinsp;0.05 for all pairwise tests, indicating that the four LLMs converged on a broadly similar physiological decision boundary, and this approach is model agnostic.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec13\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eAssociation of HAIPS with Clinical Measures of Disease Severity\\u003c/h2\\u003e \\u003cp\\u003eThe HAIPS, derived exclusively from LLM-based analysis of smartphone eye movement data, was compared with standard clinical measures of cognition, functional capacity, and motor impairment obtained during in-person examination. Among participants, higher HAIPS values were strongly associated with poorer cognitive performance on the MoCA (ρ = \\u0026minus;\\u0026thinsp;0.86, p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001), reduced functional capacity on the TFC scale (ρ = \\u0026minus;\\u0026thinsp;0.74, p\\u0026thinsp;=\\u0026thinsp;0.003), and greater motor impairment on the UHDRS motor scale (ρ\\u0026thinsp;=\\u0026thinsp;0.85, p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001) (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig5\\\" class=\\\"InternalRef\\\"\\u003e5\\u003c/span\\u003e). Demonstrating that HAIPS captures not only diagnostic status from the ocular motor data but also varies proportionally with established measures of disease severity across clinical domains.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003c/div\\u003e\"},{\"header\":\"Discussion\",\"content\":\"\\u003cp\\u003eIn this proof-of-concept study, we show that general-purpose LLMs can extract clinically meaningful information from ocular motor data acquired using a conventional smartphone application. Using eye movements features alone, the models generated a probability score, that reliably distinguished HD from healthy controls and scaled with disease severity. Our findings indicate that the models captured not only the presence of ocular motor abnormalities, but also their graded clinical relevance.\\u003c/p\\u003e \\u003cp\\u003eInitially we demonstrated clinical validity by showing strong concordance between smartphone ocular motor metrics and clinical assessments performed by trained eye movement specialists. The application generated physiologically plausible eye movement traces and summary metrics that aligned with well established clinical measures in HD (UHDRS motor score).\\u003c/p\\u003e \\u003cp\\u003eTranslating smartphone recordings into scalable and interpretable clinical inference has traditionally required expert review, curated feature extraction, or supervised ML pipelines.\\u003csup\\u003e17\\u003c/sup\\u003e These approaches are labor-intensive and depend on large, well-annotated datasets, limiting their applicability in specific domains and in rare diseases such as HD, where data scarcity is inherent.\\u003csup\\u003e18\\u003c/sup\\u003e\\u003c/p\\u003e \\u003cp\\u003eTo address this bottleneck we leveraged LLMs as general-purpose inference engines.\\u003csup\\u003e11\\u003c/sup\\u003e First, unlike conventional classifiers, LLMs do not require disease-specific training, feature reweighting, or retraining as model architectures evolve. Second, LLMs can operate directly in a zero-shot settings (e.g., without receiving any prior training examples or specific, labeled data), mapping structured, physiologically grounded eye-movement metrics to probabilistic clinical assessments without supervised optimization or model fine-tuning.\\u003c/p\\u003e \\u003cp\\u003eThe zero-shot nature of our approach is particularly important for rare neurological diseases, where large disease-specific training datasets are often unavailable.\\u003csup\\u003e9,19\\u003c/sup\\u003e By operating directly on a limited set of objective, physiologically meaningful inputs, LLMs generated clinically relevant probability estimates without retraining. Notably, all LLMs tested showed similar performance, demonstrating model-agnostic behavior and indicating that diagnostic inference was driven primarily by the structure of the ocular motor data rather than by idiosyncratic properties of any single model.\\u003csup\\u003e20\\u003c/sup\\u003e\\u003c/p\\u003e \\u003cp\\u003eConsistent with this model-agnostic behavior, all models achieved strong discrimination between HD and control participants, with AUC values exceeding 0.9. Moreover, the resulting HAIPS closely tracked clinical impairment across cognitive and motor domains (MoCA, UHDRS motor score and TFC), indicating that the models captured not only the presence of ocular motor abnormalities, but also their severity. While discrimination metrics may be influenced by the limited sample size and the comparison with healthy controls, these findings suggest that ocular motor inputs allow LLMs to unlock latent representations of normal and abnormal physiology, enabling clinically meaningful probabilistic inference via prompt-based reasoning alone.\\u003csup\\u003e21\\u003c/sup\\u003e\\u003c/p\\u003e \\u003cp\\u003eThese results are particularly timely in an era of rapid expansion of smartphone neurological assessment. Personal devices have been used to capture motor, gait, speech, and cognitive features across multiple neurodegenerative diseases.\\u003csup\\u003e22,23\\u003c/sup\\u003e However, current approaches have pitfalls, by either requiring multiple body sensors, relying on task-specific models, or depending on effortful tasks influenced by educational level or task comprehension.\\u003csup\\u003e24,25\\u003c/sup\\u003e Those facts limit scalability and increase variability in real-world settings.\\u003csup\\u003e7\\u003c/sup\\u003e\\u003c/p\\u003e \\u003cp\\u003eIn contrast, eye movements offer distinct advantages: they are rapid and easy to assess, relatively independent of patient effort and educational level, and less influenced by peripheral biomechanics. Moreover, ocular motor control is grounded in well-characterized neural circuits, providing a robust physiological substrate for neurological assessment.\\u003csup\\u003e26\\u003c/sup\\u003e We showed that when captured using consumer smartphones and interpreted through a LLM-based inference system, ocular motor signals can be transformed into clinically meaningful, probabilistic representations of neurological dysfunction (and, potentially, neurological disease diagnosis) without reliance on disease-specific training.\\u003c/p\\u003e \\u003cp\\u003eWhile this study focused on HD, its implications extend beyond a single disorder. Eye movements abnormalities are prevalent across a wide range of neurodegenerative conditions, including dementias, parkinsonian syndromes, motor neuron diseases, and cerebellar disorders.\\u003csup\\u003e26\\u003c/sup\\u003e The ability to derive disease-relevant signals from eye movements using AI approach points to a broadly applicable strategy for objective digital phenotyping across neurological diseases. Integrating ocular motor metrics with other described smartphone-derived modalities, such as motor performance, speech, and cognitive measures, may further improve sensitivity to disease burden, progression, and therapeutic response.\\u003csup\\u003e27\\u003c/sup\\u003e As disease-modifying therapies emerge for HD and other neurodegenerative disorders, scalable and objective tools for remote assessment and longitudinal monitoring are increasingly important.\\u003csup\\u003e28,29\\u003c/sup\\u003e Such digital biomarkers may ultimately serve for early detection or treatment monitoring.\\u003c/p\\u003e \\u003cp\\u003eSeveral limitations should be considered when interpreting these findings. First, the sample size was modest and derived from a single tertiary center, consistent with the exploratory nature of this study. Larger, demographically diverse cohorts will be necessary to refine estimates of diagnostic accuracy, quantify population-level variability. Second, real-world deployment will require validation across device generations, user-handling variability, as well as testing during unsupervised home use.\\u003csup\\u003e30\\u003c/sup\\u003e Third, the cross-sectional design precludes conclusions regarding sensitivity to longitudinal change, premanifest detection, or treatment effects, domains that will be critical for translation into clinical trials and disease-monitoring contexts. Finally, it remains unknown how well this approach might perform with more challenging diagnostic tasks (e.g., HD vs. Alzheimer\\u0026rsquo;s disease, as opposed to healthy controls).\\u003c/p\\u003e \\u003cp\\u003eIn conclusion, eye movement recordings acquired using a standard consumer smartphone contain physiologically informative signals that can support clinically meaningful inference. When interpreted by general-purpose LLMs under a strictly zero-shot framework, these signals yield probabilistic disease estimates that reflect both disease presence and severity. By pairing a disease-agnostic inference model with a physiological marker that reflects distributed neural circuit function, this work provides proof-of-concept for a scalable framework with broad potential across neurological diseases and other physiology-driven clinical domains.\\u003c/p\\u003e \\u003cdiv id=\\\"Sec15\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eData availability\\u003c/h2\\u003e \\u003cp\\u003eThe datasets generated and analyzed during the current study are not publicly available due to privacy and ethical restrictions related to the sharing of identifiable or potentially re-identifiable biometric data, particularly in the context of a rare genetic disorder, but are available from the corresponding author on reasonable request.\\u003c/p\\u003e \\u003c/div\\u003e\"},{\"header\":\"Declarations\",\"content\":\"\\u003ch2\\u003eCompeting Interests\\u003c/h2\\u003e\\u003cp\\u003eDavid E. Newman-Toker holds an approved patent concerning the use of the EyePhone for tracking eye and head position and has a provisional patent application for the use of the EyePhone in recording saccades and smooth pursuit. The other authors declare no competing interests.\\u003c/p\\u003e\\u003ch2\\u003eSupplementary Material\\u003c/h2\\u003e \\u003cp\\u003eSupplementary material is available at \\u003cem\\u003eBrain\\u003c/em\\u003e online\\u003c/p\\u003e\\u003ch2\\u003eEthics approval\\u003c/h2\\u003e \\u003cp\\u003e The study was approved by the Johns Hopkins Medicine Institutional Review Board (IRB00258938). All participants provided written informed consent prior to participation, and all procedures were conducted in accordance with the Declaration of Helsinki.\\u003c/p\\u003e\\u003ch2\\u003eFunding\\u003c/h2\\u003e \\u003cp\\u003eThis research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors\\u003c/p\\u003e\\u003ch2\\u003eAuthor Contribution\\u003c/h2\\u003e\\u003cp\\u003eL.A. conceived the study, designed the experimental protocol, performed the literature review, conducted data analysis, and wrote the initial draft of the manuscript. K.W. developed the smartphone application, implemented the data-processing pipeline, and contributed to the design and optimization of the ocular motor battery. D.N.T. contributed to model design and provided critical input on neurological interpretation J.B. supervised clinical aspects of the study, oversaw patient recruitment in the Huntington\\u0026rsquo;s disease clinic, and provided critical input on neurological interpretation. D.P.W.R. supervised the AI and computational framework, contributed to model design and prompt engineering, and guided analytical validation. All authors contributed to the interpretation of the data, critically revised the manuscript for important intellectual content, and approved the final version of the manuscript.\\u003c/p\\u003e\\u003ch2\\u003eAcknowledgement\\u003c/h2\\u003e\\u003cp\\u003eThe authors thank the patients and their families for their participation in this study. We also acknowledge the support of the clinical and research teams involved in the evaluation and care of individuals with Huntington\\u0026rsquo;s disease. This research did not received specific grant from any funding agency in the public, commercial, or not-for-profit sectors\\u003c/p\\u003e\\u003ch2\\u003eData Availability\\u003c/h2\\u003e\\u003cp\\u003eThe datasets generated and analyzed during the current study are not publicly available due to privacy and ethical restrictions related to the sharing of identifiable or potentially re-identifiable biometric data, particularly in the context of a rare genetic disorder, but are available from the corresponding author on reasonable request.\\u003c/p\\u003e\"},{\"header\":\"References\",\"content\":\"\\u003col\\u003e\\n\\u003cli\\u003eHuang Y, Li Y, Pan H, Han L. Global, regional, and national burden of neurological disorders in 204 countries and territories worldwide. \\u003cem\\u003eJ Glob Health\\u003c/em\\u003e. 2023;13:04160. doi:10.7189/jogh.13.04160\\u003c/li\\u003e\\n\\u003cli\\u003eBates GP, Dorsey R, Gusella JF, et al. Huntington disease. \\u003cem\\u003eNat Rev Dis Primer\\u003c/em\\u003e. 2015;1(1):15005. doi:10.1038/nrdp.2015.5\\u003c/li\\u003e\\n\\u003cli\\u003eDorsey ER, Topol EJ. State of Telehealth. Campion EW, ed. \\u003cem\\u003eN Engl J Med\\u003c/em\\u003e. 2016;375(2):154-161. doi:10.1056/NEJMra1601705\\u003c/li\\u003e\\n\\u003cli\\u003eBokinni Y. Huntington\\u0026rsquo;s disease: new gene therapy explained. \\u003cem\\u003eBMJ\\u003c/em\\u003e. 2025;390:r2029. doi:10.1136/bmj.r2029\\u003c/li\\u003e\\n\\u003cli\\u003eDagum P. Digital biomarkers of cognitive function. \\u003cem\\u003eNpj Digit Med\\u003c/em\\u003e. 2018;1(1):10. doi:10.1038/s41746-018-0018-4\\u003c/li\\u003e\\n\\u003cli\\u003eLim WS, Fan SP, Chiu SI, et al. Smartphone-derived multidomain features including voice, finger-tapping movement and gait aid early identification of Parkinson\\u0026rsquo;s disease. \\u003cem\\u003eNpj Park Dis\\u003c/em\\u003e. 2025;11(1):111. doi:10.1038/s41531-025-00953-w\\u003c/li\\u003e\\n\\u003cli\\u003eYang K, Xiong WX, Liu FT, et al. Objective and quantitative assessment of motor function in Parkinson\\u0026rsquo;s disease\\u0026mdash;from the perspective of practical applications. \\u003cem\\u003eAnn Transl Med\\u003c/em\\u003e. 2016;4(5):90-90. doi:10.21037/atm.2016.03.09\\u003c/li\\u003e\\n\\u003cli\\u003eBeam AL, Kohane IS. Big Data and Machine Learning in Health Care. \\u003cem\\u003eJAMA\\u003c/em\\u003e. 2018;319(13):1317. doi:10.1001/jama.2017.18391\\u003c/li\\u003e\\n\\u003cli\\u003eBanerjee J, Taroni JN, Allaway RJ, Prasad DV, Guinney J, Greene C. Machine learning in rare disease. \\u003cem\\u003eNat Methods\\u003c/em\\u003e. 2023;20(6):803-814. doi:10.1038/s41592-023-01886-z\\u003c/li\\u003e\\n\\u003cli\\u003eSculley D, Holt G, Golovin D, et al. Hidden Technical Debt in Machine Learning Systems. \\u003cem\\u003eNIPS\\u003c/em\\u003e. Published online January 2015:2494-2502.\\u003c/li\\u003e\\n\\u003cli\\u003eBommasani R, Hudson DA, Adeli E, et al. On the Opportunities and Risks of Foundation Models. \\u003cem\\u003earXiv\\u003c/em\\u003e. Preprint posted online 2021. doi:10.48550/ARXIV.2108.07258\\u003c/li\\u003e\\n\\u003cli\\u003eSinghal K, Azizi S, Tu T, et al. Large language models encode clinical knowledge. \\u003cem\\u003eNature\\u003c/em\\u003e. 2023;620(7972):172-180. doi:10.1038/s41586-023-06291-2\\u003c/li\\u003e\\n\\u003cli\\u003eLeigh RJ, Zee DS. \\u003cem\\u003eThe Neurology of Eye Movements\\u003c/em\\u003e. 5th ed. Oxford university press; 2015.\\u003c/li\\u003e\\n\\u003cli\\u003e\\u003cem\\u003eEye Tracking: A Comprehensive Guide to Methods and Measures\\u003c/em\\u003e. Oxford university press; 2011.\\u003c/li\\u003e\\n\\u003cli\\u003eUnified Huntington\\u0026rsquo;s disease rating scale: Reliability and consistency. \\u003cem\\u003eMov Disord\\u003c/em\\u003e. 1996;11(2):136-142. doi:10.1002/mds.870110204\\u003c/li\\u003e\\n\\u003cli\\u003eWei J, Bosma M, Zhao VY, et al. Finetuned Language Models Are Zero-Shot Learners. \\u003cem\\u003earXiv\\u003c/em\\u003e. Preprint posted online 2021. doi:10.48550/ARXIV.2109.01652\\u003c/li\\u003e\\n\\u003cli\\u003eEllis RJ, Sander RM, Limon A. Twelve key challenges in medical machine learning and solutions. \\u003cem\\u003eIntell-Based Med\\u003c/em\\u003e. 2022;6:100068. doi:10.1016/j.ibmed.2022.100068\\u003c/li\\u003e\\n\\u003cli\\u003eNgiam KY, Khor IW. Big data and machine learning algorithms for health-care delivery. \\u003cem\\u003eLancet Oncol\\u003c/em\\u003e. 2019;20(5):e262-e273. doi:10.1016/S1470-2045(19)30149-4\\u003c/li\\u003e\\n\\u003cli\\u003eDecherchi S, Pedrini E, Mordenti M, Cavalli A, Sangiorgi L. Opportunities and Challenges for Machine Learning in Rare Diseases. \\u003cem\\u003eFront Med\\u003c/em\\u003e. 2021;8:747612. doi:10.3389/fmed.2021.747612\\u003c/li\\u003e\\n\\u003cli\\u003eLiang P, Bommasani R, Lee T, et al. Holistic Evaluation of Language Models. Published online 2022. doi:10.48550/ARXIV.2211.09110\\u003c/li\\u003e\\n\\u003cli\\u003eLake BM, Ullman TD, Tenenbaum JB, Gershman SJ. Building machines that learn and think like people. \\u003cem\\u003eBehav Brain Sci\\u003c/em\\u003e. 2017;40:e253. doi:10.1017/S0140525X16001837\\u003c/li\\u003e\\n\\u003cli\\u003eBot BM, Suver C, Neto EC, et al. The mPower study, Parkinson disease mobile data collected using ResearchKit. \\u003cem\\u003eSci Data\\u003c/em\\u003e. 2016;3(1):160011. doi:10.1038/sdata.2016.11\\u003c/li\\u003e\\n\\u003cli\\u003eKoerner J, Zou E, Karl JA, et al. Towards scalable screening for the early detection of Parkinson\\u0026rsquo;s disease: validation of an iPad-based eye movement assessment system against a clinical-grade eye tracker. \\u003cem\\u003eNpj Park Dis\\u003c/em\\u003e. 2025;11(1):233. doi:10.1038/s41531-025-01079-9\\u003c/li\\u003e\\n\\u003cli\\u003eMeng X, D\\u0026rsquo;Arcy C. Education and Dementia in the Context of the Cognitive Reserve Hypothesis: A Systematic Review with Meta-Analyses and Qualitative Analyses. Laks J, ed. \\u003cem\\u003ePLoS ONE\\u003c/em\\u003e. 2012;7(6):e38268. doi:10.1371/journal.pone.0038268\\u003c/li\\u003e\\n\\u003cli\\u003eDel Din S, Elshehabi M, Galna B, et al. Gait analysis with wearables predicts conversion to Parkinson disease. \\u003cem\\u003eAnn Neurol\\u003c/em\\u003e. 2019;86(3):357-367. doi:10.1002/ana.25548\\u003c/li\\u003e\\n\\u003cli\\u003eAntoniades CA, Kennard C. Ocular motor abnormalities in neurodegenerative disorders. \\u003cem\\u003eEye\\u003c/em\\u003e. 2015;29(2):200-207. doi:10.1038/eye.2014.276\\u003c/li\\u003e\\n\\u003cli\\u003eSun Y meng, Wang Z yun, Liang Y yuan, Hao C wei, Shi C he. Digital biomarkers for precision diagnosis and monitoring in Parkinson\\u0026rsquo;s disease. \\u003cem\\u003eNpj Digit Med\\u003c/em\\u003e. 2024;7(1):218. doi:10.1038/s41746-024-01217-2\\u003c/li\\u003e\\n\\u003cli\\u003eeClinicalMedicine. A new hope for patients with Huntington\\u0026rsquo;s disease? \\u003cem\\u003eeClinicalMedicine\\u003c/em\\u003e. 2025;88:103612. doi:10.1016/j.eclinm.2025.103612\\u003c/li\\u003e\\n\\u003cli\\u003eCummings J, Lee G, Nahed P, et al. Alzheimer\\u0026rsquo;s disease drug development pipeline: 2022. \\u003cem\\u003eAlzheimers Dement Transl Res Clin Interv\\u003c/em\\u003e. 2022;8(1):e12295. doi:10.1002/trc2.12295\\u003c/li\\u003e\\n\\u003cli\\u003eGoetz L, Seedat N, Vandersluis R, Van Der Schaar M. Generalization\\u0026mdash;a key challenge for responsible AI in patient-facing clinical applications. \\u003cem\\u003eNpj Digit Med\\u003c/em\\u003e. 2024;7(1):126. doi:10.1038/s41746-024-01127-3\\u003c/li\\u003e\\n\\u003c/ol\\u003e\"}],\"fulltextSource\":\"\",\"fullText\":\"\",\"funders\":[],\"hasAdminPriorityOnWorkflow\":false,\"hasManuscriptDocX\":true,\"hasOptedInToPreprint\":true,\"hasPassedJournalQc\":\"\",\"hasAnyPriority\":false,\"hideJournal\":true,\"highlight\":\"\",\"institution\":\"\",\"isAcceptedByJournal\":false,\"isAuthorSuppliedPdf\":false,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":false,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"researchsquare\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":true,\"externalIdentity\":\"\",\"sideBox\":\"\",\"snPcode\":\"\",\"submissionUrl\":\"/submission\",\"title\":\"Research Square\",\"twitterHandle\":\"researchsquare\",\"acdcEnabled\":true,\"dfaEnabled\":false,\"editorialSystem\":\"\",\"reportingPortfolio\":\"\",\"inReviewEnabled\":false,\"inReviewRevisionsEnabled\":true},\"keywords\":\"Mobile Health, Digital Biomarkers, Huntington’s disease; eye tracking;; artificial intelligence, large language models\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-8844897/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-8844897/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"\\u003cp\\u003eDigital biomarkers derived from consumer devices offer new opportunities for remote neurological assessment. However, most AI-based approaches depend on large, disease-specific training datasets, limiting their applicability in rare disorders. Large language models (LLMs), trained on broad medical corpora, may enable clinically meaningful inference without disease-specific model training when provided with structured physiological inputs.\\u003c/p\\u003e \\u003cp\\u003eIn this prospective proof-of-concept study, individuals with genetically confirmed Huntington\\u0026rsquo;s disease (HD) and age-matched healthy controls completed an ocular motor assessment using an in-house-developed smartphone application. Quantitative eye movement metrics were validated against expert neurologist ratings and subsequently provided to LLMs using a structured prompt. Models generated an AI-assigned HD probability score (HAIPS) based exclusively on ocular motor data.\\u003c/p\\u003e \\u003cp\\u003eTwenty-six participants were included. Smartphone-derived metrics showed strong agreement with clinical ratings (Spearman ρ 0.76\\u0026ndash;0.95; all p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.001). HAIPS reliably discriminated individuals with HD from controls (AUC 0.879\\u0026ndash;0.944), with no significant differences across models. Among HD participants, higher HAIPS correlated with established motor and cognitive measures (Spearman ρ 0.74\\u0026ndash;0.86; all p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.01).\\u003c/p\\u003e \\u003cp\\u003eThese findings demonstrate that LLMs can generate clinically meaningful probabilistic assessments of HD from smartphone-derived ocular motor data without disease-specific training, highlighting a scalable framework for AI-supported assessment in neurodegenerative disorders.\\u003c/p\\u003e\",\"manuscriptTitle\":\"Smartphone-Derived Ocular Motor Biomarkers Enable AI to Assess Neurodegeneration in Huntington’s Disease\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2026-03-08 14:36:57\",\"doi\":\"10.21203/rs.3.rs-8844897/v1\",\"editorialEvents\":[{\"type\":\"communityComments\",\"content\":0}],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"researchsquare\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":true,\"externalIdentity\":\"\",\"sideBox\":\"\",\"snPcode\":\"\",\"submissionUrl\":\"/submission\",\"title\":\"Research Square\",\"twitterHandle\":\"researchsquare\",\"acdcEnabled\":true,\"dfaEnabled\":false,\"editorialSystem\":\"\",\"reportingPortfolio\":\"\",\"inReviewEnabled\":false,\"inReviewRevisionsEnabled\":true}}],\"origin\":\"\",\"ownerIdentity\":\"65c589ab-6e83-4cc1-8047-6407a780d64b\",\"owner\":[],\"postedDate\":\"March 8th, 2026\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"posted\",\"subjectAreas\":[],\"tags\":[],\"updatedAt\":\"2026-03-26T19:54:41+00:00\",\"versionOfRecord\":[],\"versionCreatedAt\":\"2026-03-08 14:36:57\",\"video\":\"\",\"vorDoi\":\"\",\"vorDoiUrl\":\"\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-8844897\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-8844897\",\"identity\":\"rs-8844897\",\"version\":[\"v1\"]},\"buildId\":\"XKTyCvWXoU3ODBz1xrDgd\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}