Public defence: Pedro Maria Rego Lencastre E Silva

Eye-tracking measures where we look and how our pupils change, giving a quick, non-invasive window into attention, perception, and mental effort. Because it’s becoming easier to do, inside VR headsets and even with ordinary cameras, it has real potential as an early screening tool for conditions like autism and ADHD, where earlier support can improve long-term outcomes.

But research on eye movements is oddly fragmented. Physicists often model gaze as a simple kind of random motion; clinicians use standardized tests and split behaviour into two phases: fixations (when we take in information) and saccades (fast jumps between points); and AI can learn directly from complex data, but the result is often a “black box” that’s hard to trust or interpret in healthcare.

In this thesis, we connect these approaches and show that simple, interpretable models that switch between a few states can reproduce key patterns of gaze better than the generative AI models we tested. 

Building on that, we show how these models, combined with standard machine-learning classifiers, can efficiently detect ADHD-related gaze signatures. We also present evidence that gaze is best described by models that explicitly represent switching between phases, and we report new findings suggesting 

• tiny rhythmic movements during fixations relate to individual limits in temporal perception
• saccades contain distinct subtypes, hinting that we may need a more detailed taxonomy of “what counts” as an eye movement

Overall, we aim to give a more unified picture of how gaze behaves, while making it more useful for practical, trustworthy applications in intelligent health.