Abstract
Eye movements are disrupted in many neurodegenerative diseases and are frequent and early features in conditions affecting the cerebellum. Characterizing eye movements is important for diagnosis and may be useful for tracking disease progression and response to therapies. Assessments are limited as they require an in-person evaluation by a neurology subspecialist or specialized and expensive equipment. We tested the hypothesis that important eye movement abnormalities in cerebellar disorders (i.e., ataxias) could be captured from iPhone video. Videos of the face were collected from individuals with ataxia (n = 102) and from a comparative population (Parkinson's disease or healthy participants, n = 61). Computer vision algorithms were used to track the position of the eye which was transformed into high temporal resolution spectral features. Machine learning models trained on eye movement features were able to identify abnormalities in smooth pursuit (a key eye behavior) and accurately distinguish individuals with abnormal pursuit from controls (sensitivity = 0.84, specificity = 0.77). A novel machine learning approach generated severity estimates that correlated well with the clinician scores. We demonstrate the feasibility of capturing eye movement information using an inexpensive and widely accessible technology. This may be a useful approach for disease screening and for measuring severity in clinical trials.
Introduction
The cerebellum acts as the brain's motor control center, and one of its earliest functions to degrade in neurodegenerative diseases is smooth pursuit—the ability of the eyes to track a moving object fluidly. When the cerebellum is impaired, as seen in Spinocerebellar Ataxias (SCAs), this fluid tracking breaks down into "saccadic pursuit," characterized by small, jerky "catch-up" movements.
Diagnosing these abnormalities currently requires highly trained neurologists or expensive, specialized infrared eye-tracking equipment. This creates significant barriers for patients in remote areas and complicates frequent monitoring in clinical trials. This research demonstrates a method to bridge this gap by transforming a consumer smartphone into a clinical-grade diagnostic tool for cerebellar ataxia.
Methodology
The researchers implemented a two-device recording setup: an iPad Pro displayed a moving dot stimulus while an iPhone 8+ recorded the participant's face at 240 frames per second (fps). This high frame rate is critical for capturing micro-movements that standard 30fps or 60fps video cannot resolve.
The technical pipeline involves four primary stages:
- Computer Vision: The Intraface algorithm tracked 12 facial landmarks and the iris center in every frame.
- Signal Cleaning: Blinks were detected via eye aspect ratios and corrected using cubic interpolation, followed by low-pass and median filtering to remove noise.
- Spectral Analysis: Using ConceFT (Concentrated Frequency and Time), a multitapered synchrosqueezed transform, the team extracted high-resolution time-frequency features from the eye position data.
- Machine Learning: A linear Support Vector Machine (SVM) was trained to classify pursuit as "abnormal" or "typical." Severity was graded using a pairwise comparison approach with L1-regularized logistic regression.
Clinical Dataset
The study involved a diverse cohort of 163 participants, including 102 patients with various ataxia subtypes (SCA-1, 2, 3, 6, Friedreich’s Ataxia, etc.), 43 Parkinson’s Disease (PD) patients, and 18 healthy controls.
The "Typical" group for classification purposes included healthy controls, PD patients, and ataxia patients with a clinical oculomotor score of zero. As shown in the Venn diagram, saccadic pursuit was a dominant feature, often overlapping with nystagmus and dysmetric saccades.
Key Insights and Results
- The 1.5–2.5 Hz "Signature": A primary discovery was that saccadic pursuit manifests as a specific power spike in the 1.5–2.5 Hz frequency band. This represents the periodicity of the corrective "catch-up" jumps made by the eyes.
- Classification Performance: The model achieved an AUC of 0.85, with a sensitivity of 0.84 and specificity of 0.77 in distinguishing abnormal pursuit from typical tracking.
- Clinical Correlation: Machine learning severity estimates showed a Pearson correlation of 0.63 with the Brief Ataxia Rating Scale (BARS) oculomotor subscore, validating the model against clinician observations.
- Specificity to Ataxia: The features were highly specific to cerebellar tracking dysfunction, as the model was significantly better at identifying ataxia-related issues than simply distinguishing ataxia from Parkinson's disease.
Discussion and Future Directions
While the results are promising, approximately 14% of participants were excluded due to excessive movement or technical issues, suggesting a need for real-time user feedback in future applications. Additionally, the current setup requires two devices; future iterations could leverage high-speed front-facing cameras on newer smartphones to consolidate the tool into a single-device app.
The next critical step is longitudinal tracking—proving the tool can sensitively detect disease progression or response to therapy in a single patient over time. This would establish the smartphone as a vital instrument for both clinical practice and decentralized clinical trials.
Summary updated