Deep learning-based eye movement autism screening model construction method and related device
By constructing a multi-scale one-dimensional convolutional neural network and a two-layer long short-term memory network, and combining transient response and continuous state attention pathways, the problem that existing models cannot distinguish eye movement behavior patterns is solved, thus improving the accuracy and interpretability of autism screening.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAMEN UNIV TAN KAH KEE COLLEGE
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-17
AI Technical Summary
Existing CNN-LSTM-Attention eye-tracking screening models employ a single global attention mechanism, which cannot distinguish between transient reactive attention and sustained state attention in eye-tracking behavior. This results in an inability to accurately capture transient reaction abnormalities and sustained preference features related to autism, limiting the improvement of screening accuracy and the clinical interpretability of model output.
A multi-scale one-dimensional convolutional neural network is constructed to extract local spatiotemporal features. Temporal dependencies are modeled through a two-layer long short-term memory network. The hidden state sequence is input into the transient response attention path and the persistent state attention path. Attention weights are calculated respectively to generate a global representation vector. Finally, the classification probability is output through a fully connected classification layer.
By separating transient reactive attention and sustained state attention, the model's recognition accuracy of core behavioral markers of autism was significantly improved, achieving a more complete and three-dimensional characterization of eye-movement behavior and enhancing the model's clinical interpretability.
Smart Images
Figure CN122417399A_ABST