A Data-Driven Unsupervised Framework for Discovering Interpretable Gaze-Based Behavioral Pseudo-Zones in Children...

Area of Science:

• Neuroscience
• Computer Science
• Developmental Psychology

Background:

• Children with autism spectrum disorder (ASD) display atypical attention regulation and visual behaviors.
• Existing ASD eye-tracking datasets often lack detailed annotations, hindering supervised learning.
• This limits the direct analysis of moment-to-moment behaviors in ASD.

Purpose of the Study:

• To develop an interpretable, gaze-based unsupervised framework for discovering behavioral pseudo-zones in unlabeled ASD eye-tracking data.
• To address the limitations of supervised learning due to missing annotations in ASD datasets.
• To enable data-driven analysis of visual behavior in autism.

Main Methods:

• Raw gaze recordings from ASD participants were segmented into temporal windows.
• Interpretable gaze features (e.g., dispersion, fixation duration, pupil size) were extracted.
• Clustering models (K-means, GMM, Agglomerative, HDBSCAN) and window sizes were compared.

Main Results:

• A configuration using 1000 ms windows and K-means clustering (k=4) yielded an interpretable four-zone structure.
• These four behavioral pseudo-zones showed statistically significant differences across all gaze features (p < 0.05).
• Principal Component Analysis (PCA) supported the distinctness of these zones, explaining 72.3% of variance.

Conclusions:

• Unlabeled ASD gaze data can be effectively organized into interpretable behavioral pseudo-zones.
• The proposed unsupervised framework is transparent and feature-based.
• This work provides a foundation for future gaze-based behavioral analysis and AI research in autism.