A study on eye movement trajectory classification for strabismus screening via integration of global dependencies and temporal dynamics.

Strabismus is a binocular visual disorder characterized by deviation of the visual axes, which may lead to irreversible impairments such as amblyopia if not recognized and intervened in time. To address the limitations of existing methods-including reliance on costly hardware, limited adaptability to dynamic conditions, and insufficient capability in strabismus subtype classification-this study proposes a wearable eye-tracker-based deep learning framework for categorizing eye movement sequences into controls, exotropes, and esotropes. The model integrates spatial, temporal, and global features using a hierarchical structure: spatial features are extracted by a modified ResNet-18, sequence dynamics are modeled using a bidirectional LSTM, and long-range temporal dependencies are captured in combination with a lightweight Transformer encoder. A conditional trigger-based model filtering strategy is introduced to enhance the model's sensitivity to minority-class samples. Experimental results demonstrate that the proposed method achieves an overall sample-level classification accuracy of 93.3%, with particularly strong performance in identifying adolescent esotropes, validating its potential for clinical and primary screening applications.