TS-EyeTrack: Learning Gaze-Derived Biomarkers for Autism Spectrum Disorder Classification

Introduction: The early and objective identification of Autism Spectrum Disorder (ASD) continuously relies on quantitative behavioral biomarkers such as gaze dynamics. This study developed an interpretable machine learning framework that uses high-resolution eye-tracking data for ASD classification and stimulus-specific gaze behavior decoding. Methods: A dataset that consists of 184 samples from 25 ASD participants was processed to extract temporal-spatial features, including the movement velocity of eyes, Point-Of-Regard (POR) deviation, fixation duration asymmetry, gaze discrepancy, and stimulus-wise tracking ratio. Principal Component Analysis (PCA) was used for dimensionality reduction and visualization, while Kmeans clustering identified latent behavioral subtypes. Supervised modeling employed Random Forest for ASD vs. TD classification and Gradient Boosting for eight stimulus categories. The performance of the model was evaluated through a cross-validation, an ablation analysis, and Leave- One-Participant-Out (LOPO) testing. Temporal autocorrelation and fixation-decay analyses assessed gaze rhythm alterations. Results: The Random Forest model achieved 98.5% accuracy (F1-score: 98.3%) for ASD classification, and the Gradient Boosting Classifier achieved 97% accuracy across eight cognitive stimulus classes. PCA and clustering revealed distinct attentional subtypes among participants. Discussion: Findings show that meaningful differences in gaze patterns, attentional stability, and temporal gaze rhythms in ASD participants demonstrate the utility of gaze-derived biomarkers for behavioral profiling. Conclusion: The study’s novelty lies in integrating high-resolution gaze biomarkers with a two-stage interpretable machine learning pipeline, enabling fine-grained stimulus-level behavioral decoding and clustering-based ASD subtyping.