Towards Affordable Smartphone Eye Tracking for Nystagmus Analysis and Monitoring

Nystagmus, an involuntary oscillatory eye movement, is an important diagnostic marker for various neurological and vestibular disorders. However, current clinical eye tracking systems are expensive, require specialized equipment and training, thus limiting their usability at scale. This study proposes a smartphone-based eye tracking pipeline for nystagmus analysis, utilizing a deep learning-based segmentation model and a circle-fitting algorithm to estimate pupil position. Two variations of U-Net segmentation models - a 3-class (sclera, iris, background) and a 4-class (sclera, iris, pupil, background) U-Net - were trained on publicly available MOBIUS dataset. Then applied and evaluated on 12 optokinetic nystagmus (OKN) videos with varying gaze positions. On the test set the 3-class model performed best across cross entropy loss, intersection over union loss, and DICE score metrics (0.051, 0.081, and 0.956 respectively), compared to the 4-class model's 0.068, 0.162, and 0.903 respectively. On the tracking task, when evaluated against ground truth tracking measured using the ICS Impulse, both methods were able to capture the nystagmus movements reasonably well (avg R² of 0.755 and 0.768 for 3 and 4-class methods across 12 OKN tests). No statistical difference in the performance of the 3-class model compared to the 4-class model on R² (p-value 0.80). These findings highlight the potential of smartphone-based eye tracking as a cost-effective tool for objective nystagmus assessment, with further refinements needed to improve tracking performance. This approach could enhance diagnostic accuracy and accessibility in clinical settings.