Improving Autism Diagnosis Across Ages Using Eye-Tracking and Temporal Transformer Models

Abstract

Variation in gaze behavior due to age is currently a considerable challenge in building reliable eye-tracking systems for Autism Spectrum Disorder (ASD) diagnosis. However, existing strategies often focus on static gaze representation or dataset-based information, which can lead to limited generalization of findings depending on developmental groups and heterogeneous recording conditions. In this paper, we present a temporal transformer-based system for ASD classification using eye-tracking sequences. This allows you to model gaze behavior as a structured temporal process in the context of contextual attention, as well as employing entropy-based modeling for various distributions of variability over time and temporal consistency constraints to capture sequential gaze dynamics related to ASD behavioral patterns. The framework was evaluated using public eye-tracking corpus containing temporally ordered gaze recordings from ASD and TD participants across age groups. Five sequential experiments on baseline classification, class-balancing analysis, cross-age evaluation, ablation analysis, and cross-dataset transfer learning were performed to conduct experiment-based evaluations. Model performed 0.91 in in-domain Area Under the Receiver Operating Characteristic Curve (AUC) and 0.81 in F1-score on the primary eye-tracking dataset. In the cross-dataset assessment stage, the framework presented a relatively stable performance, with an AUC of 0.85 and an average F1-score of 0.74, irrespective of differences in participant distributions and recording conditions. Ablation analysis also revealed that entropy regularization and temporal consistency mechanisms played a significant role in model stability and classification performance. The ablation analysis provides additional insight into the contribution of the proposed framework components beyond the overall classification performance. Removing the entropy-based regularization reduced the model’s ability to represent variability in gaze allocation, whereas removing the temporal-consistency regularization resulted in less stable sequence representations during learning. These observations indicate that the proposed components complement the transformer-based sequence encoder by improving representation stability and preserving diagnostically relevant temporal information. Rather than acting as independent classifiers, the regularization mechanisms serve as supporting constraints that enhance the quality and robustness of the learned temporal representations. The results indicate that temporally structured gaze modeling is more robust, interpretable, and general in comparison to static gaze representations. In summary, the presented framework can represent a scalable and developmentally appropriate approach to gaze-based ASD classification and support the implementation of trusted neurodevelopmental screening systems.