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

Problem — The paper addresses the lack of robust AI-enabled video-oculographic solutions for detecting saccadic signatures in neurological diseases, which are critical for screening and localizing brain abnormalities. Current methods face challenges due to privacy concerns and limited datasets, making it difficult to develop effective digital biomarkers. This work is a preprint and has not undergone peer review.

Method — The authors propose GenEyePose, a fully synthetic, patient-free pipeline that generates multimodal eye movement data for saccade analysis. The pipeline utilizes a generative approach to create diverse saccadic patterns, which are then used to train a deep learning classifier. The classifier is designed to differentiate between normal saccadic movements and those indicative of hypometria and hypermetria. The architecture specifics, including the type of neural network employed and the loss function used for training, are not disclosed in the abstract. The training process leverages the synthetic dataset to enhance the model’s generalizability, and the evaluation is conducted on real-world clinical data.

Results — The model achieved an Area Under the Receiver Operating Characteristic (AUROC) score of 0.76 and a sensitivity of 0.71 when tested on clinical data. These results indicate that the synthetic data generated by GenEyePose can effectively generalize to real-world applications, outperforming traditional methods that rely on limited patient data. The performance metrics suggest that the model could serve as a viable screening tool in both at-home and emergency room settings.

Limitations — The authors acknowledge that the synthetic nature of the dataset may not capture all the complexities of real-world eye movements, potentially limiting the model’s applicability in certain clinical scenarios. Additionally, the performance metrics, while promising, indicate that there is still room for improvement in sensitivity and specificity. The lack of detailed architecture and training compute information also limits reproducibility and understanding of the model’s capabilities.

Why it matters — The development of GenEyePose has significant implications for the field of digital neurophysiology, as it provides a novel approach to generating eye movement data without the need for patient involvement. This could facilitate the rapid development of screening tools for neurological conditions, making them more accessible and cost-effective. The findings underscore the potential of synthetic data in training AI models for clinical applications, paving the way for future research in this area. As published in arXiv cs.CV, this work could inspire further exploration into synthetic data generation techniques for other biomarker development efforts.