Autism Detection Based on Eye Movement Sequences on the Web: A Scanpath Trend Analysis Approach
Sukru Eraslan, Yeliz Yesilada, Victoria Yaneva, Simon Harper · 2020 · Proceedings of the 17th International Web for All Conference (W4A) · doi:10.1145/3371300.3383340
Summary
This paper investigates whether sequential eye-movement data — the order in which people look at elements on web pages — can be used to detect autism, improving upon the authors' previous non-sequential approach that achieved 75% accuracy but was unstable across different web pages. The authors use Scanpath Trend Analysis (STA), an algorithm designed to identify a representative "trending path" from a group of users' eye movements on a web page. The approach works by first computing separate trending paths for people with autism and neurotypical people using training data, then classifying new individuals by measuring how similar their personal scanpath is to each group's trending path using Levenshtein (string-edit) distance. The STA algorithm operates in three stages: a preliminary stage that maps fixations to visual elements on the page, a first pass that identifies trending element instances shared by a threshold percentage of participants, and a second pass that combines these instances into a final trending path ordered by sequential priority values. The key motivation is that previous eye-tracking research has consistently shown that people with autism exhibit different visual attention patterns — more fixations, more transitions between page elements, more frequent fixations on task-irrelevant elements, and comparatively shorter fixation durations — and these differences should manifest in the sequential order of attention, not just in aggregate statistics. The study used an existing eye-tracking dataset of 15 verbal, highly-independent adults with autism and 15 neurotypical adults viewing six popular websites (Apple, Babylon, AVG, Yahoo, GoDaddy, BBC) during both browsing and searching tasks.
Key findings
The STA approach achieved approximately 60% mean accuracy for both browsing and searching tasks across individual web pages, with notably more stable results than the previous non-sequential logistic regression approach. For browsing tasks, the STA approach achieved a mean accuracy of 60.00% (SD: 3.2) compared to 54.67% (SD: 6.15) for logistic regression. For searching tasks, STA achieved 58.20% (SD: 3.12) versus 56.00% (SD: 10.84). While the absolute accuracy of STA was sometimes lower than logistic regression on individual pages, the critical advantage was consistency — the standard deviation of accuracy across pages was roughly halved, meaning the approach is not dependent on specific web pages to produce reliable results. The trending paths themselves revealed qualitative differences: the autism group's trending paths contained more repeated elements and more transitions between elements, consistent with the detail-focused cognitive style associated with autism. The tolerance level parameter of the STA algorithm (ranging from 0.01 to 1.00) had minimal effect on overall results, with standard deviations of F1-scores around 0.05 across all tolerance levels. The authors note that these results were achieved with verbal, highly-independent adults with autism and cannot be generalized to individuals who are non-verbal, have intellectual disability, or do not frequently use the web.
Relevance
This research contributes to the development of objective, technology-based screening tools for autism, addressing the current diagnostic process that is subjective, expensive, and resource-intensive — with demand in the UK alone far exceeding diagnostic capacity. For accessibility practitioners, the findings have dual significance. First, the demonstrated differences in how people with autism sequentially attend to web page elements provides empirical evidence that can inform autism-friendly web design — understanding which elements attract attention and in what order could guide layout decisions. Second, the approach of using everyday web browsing behaviour as a basis for screening is far less intrusive and expensive than alternatives like fMRI (79-86% accuracy) or EEG (94% accuracy), which require specialized clinical settings. The observation that sequential analysis provides more stable cross-page results is particularly important for practical screening tools, as it means participants would need to view fewer pages to receive a reliable classification. The work also has implications for detecting other attention-related conditions such as dyslexia through similar eye-tracking approaches.
Tags: autism · eye tracking · machine learning · web accessibility · scanpath analysis · screening · neurodevelopmental disorders
Standards referenced: DSM-5