Detecting Autism Based on Eye-Tracking Data from Web Searching Tasks
Victoria Yaneva, Le An Ha, Sukru Eraslan, Yeliz Yesilada, Ruslan Mitkov · 2018 · Proceedings of the 15th International Web for All Conference (W4A) · doi:10.1145/3192714.3192819
Summary
This paper investigates whether eye-tracking data collected during everyday web tasks can be used to distinguish between people with and without autism, potentially enabling a low-cost, accessible screening approach. The study collected gaze data from 15 adults with clinically diagnosed high-functioning autism or Asperger's syndrome and 15 neurotypical controls as they performed two types of tasks on six web pages of varying visual complexity: browsing (free exploration for up to two minutes) and searching (locating specific information within 30 seconds). The rationale is grounded in established differences in visual attention between autistic and neurotypical individuals, including the Weak Central Coherence Theory (which posits that autistic cognition favors local detail processing over global context) and stimulus overselectivity (tunnel vision-like focus on one sensory modality). The authors used logistic regression with 100-fold cross-validation, where each fold trained on 10 participants per group and tested on the remaining 5, ensuring no participant's data appeared in both training and test sets. Gaze features included time to first view, total viewing time, number of fixations, and revisits per area of interest (AOI), alongside non-gaze features like page visual complexity, participant gender, and whether an AOI contained the correct search answer. The study also compared page-specific AOIs (defined by the VIPS page segmentation algorithm) against generic 2x2 grid AOIs.
Key findings
The best-performing classifier achieved 0.75 accuracy (95% CI: 0.73-0.78) for the Search task using gaze features on selected web pages (Apple + Babylon), and 0.71 accuracy (95% CI: 0.69-0.73) for the Browse task (Apple + AVG). The Search task consistently outperformed Browse, likely because searching for specific information elicits larger between-group differences in visual attention patterns. Confusion matrices showed balanced prediction across both classes (ASD: 0.808, Control: 0.698 for the best Search classifier). Notably, page visual complexity and participant gender did not significantly affect classification performance. Page-specific AOIs outperformed generic 2x2 grid AOIs for the Search task (0.75 vs 0.56) but performed similarly for Browse (0.71 vs 0.70), suggesting that meaningful page segmentation matters most when the task requires locating specific content. The selection of which web pages to include significantly affected results — certain pages elicited greater between-group differences than others. All gaze data, R code, visual stimuli, and task descriptions were made freely available for replication.
Relevance
This is the first study to use eye-tracking gaze data specifically for autism detection, opening a novel research direction that connects web accessibility research with clinical screening. The practical implications are substantial: current ASD diagnosis is long, expensive, subjective, and inaccessible in many countries, while web-based behavioral data could be collected cheaply at scale. The authors envision developing serious games for autism screening based on web-interaction patterns, which could be especially valuable in low- and middle-income countries with limited clinical resources. For web accessibility practitioners, the study provides concrete evidence that autistic users process web content differently — exhibiting shorter fixations, longer scan paths, more transitions between elements, and more fixations on irrelevant content — which reinforces the need for simpler, well-structured page layouts. The authors appropriately flag ethical considerations around behavioral profiling and note that this approach is meant to supplement, not replace, clinical diagnosis. A key limitation is the small sample size (15 per group), though the rigorous cross-validation approach helps mitigate this.
Tags: autism · eye tracking · machine learning · web accessibility · cognitive accessibility · screening · diagnostic classification · visual attention · user study · gaze data · web searching · open data · neurodivergence