Abstract
Blepharospasm is a focal dystonia characterized by involuntary eyelid contractions that impair vision and social function. The subtle clinical signs of blepharospasm make early and accurate diagnosis difficult, delaying timely intervention. In this study, we propose a dual cross-attention deep learning framework that integrates temporal video features and facial landmark dynamics to assess blepharospasm severity, frequency, and diagnosis from smartphone-recorded facial videos. A retrospective dataset of 847 patient videos collected from two hospitals (2016-2023) was used for model development. The model achieved high accuracy for severity (0.828) and frequency (0.82), and moderate performance for diagnosis (0.674).SHAP analysis identified case-specific video fragments contributing to predictions, enhancing interpretability. In a prospective evaluation on an independent dataset (N = 179), AI assistance improved junior ophthalmologist's diagnostic accuracy by up to 18.5%. These findings demonstrate the potential of an explainable, smartphone-compatible video model to support early detection and assessment of blepharospasm.