Szczepaniak, Dominik (2026) Shaping and evaluating cognitive training in virtual reality using artificial intelligence agents. PhD thesis, University of Glasgow.
Full text available as:|
PDF
Download (10MB) |
Abstract
Cognitive training that targets sustained attention and working memory is useful in education, clinical rehabilitation and interactive human-technology systems, but its benefit depends on keeping task difficulty matched to the person. Virtual reality (VR) offers immersion and settings representative of everyday tasks, and allows difficulty to be changed in real time. Existing adaptive systems, however, often rely on simple performance heuristics, affect or estimate visual complexity rather than task-related cognitive load. Moreover, cognitive load is usually judged against the participant’s own ratings, which are themselves an imperfect reference. The central hypothesis of this thesis is that eyetracking and peripheral physiological signals recorded in VR can estimate how difficult a task feels to the person doing it, well enough for a model to adjust difficulty in real time with outcomes comparable to participant-driven control, and that the estimation part holds for older adults once age-related differences in the signal are taken into account. We developed a bespoke VR sustained attention task and collected eye-tracking, photoplethysmography and galvanic skin response data from 54 younger (aged 18-58) and 19 older (aged 60+) adults at fixed difficulty levels. Models evaluated on unseen participants performed above chance (younger adults: accuracy 0.56, AUC 0.72), with eye-tracking carrying most of the signal and physiological features alone close to chance. Performance was lower in older adults (accuracy 0.49, AUC 0.66), and the most informative features appeared to differ between age groups. A bidirectional LSTM with self-attention reached a slightly higher macro-F1 than gradient-boosted trees on the same data (0.50 against 0.47), and performed better than a model given raw signals only.
We then deployed the LSTM in real time with 54 younger adults (51 analysed), comparing model-driven and participant-driven difficulty in single- and dual-task conditions. Per-trial agreement with participants’ ratings was low (accuracy 36.7%, AUC 0.63), yet the two forms of control appeared to produce comparable outcomes. There were no significant differences in difficulty-weighted game performance, arithmetic errors or subjective workload, including frustration. Final difficulty was equivalent in single-task conditions and slightly higher under model control in dual-task conditions, a difference that was borderline after correction. Just Right ratings were more frequent in the real-time cohort than in the offline cohort (57% and 39%).
Finally, we compared subjective ratings with performance-derived reference measures. Agreement between the two appeared fair and slightly lower in real time, where the steps between difficulty levels are less discrete. The same features seemed to predict performance-derived labels more accurately than subjective ratings in every matched comparison. This suggests that self-report may be a less stable reference for real-time adaptation.
Within a single session with younger adults, model-driven adaptation appears to have produced outcomes comparable to participant-driven control. The dataset (EyeDAPT) and the experimental software are publicly available.
| Item Type: | Thesis (PhD) |
|---|---|
| Qualification Level: | Doctoral |
| Subjects: | Q Science > QA Mathematics > QA75 Electronic computers. Computer science R Medicine > RC Internal medicine > RC0321 Neuroscience. Biological psychiatry. Neuropsychiatry |
| Colleges/Schools: | College of Science and Engineering > School of Computing Science |
| Funder's Name: | Engineering and Physical Sciences Research Council (EPSRC) |
| Supervisor's Name: | Deligianni, Professor Fani and Harvey, Professor Monika |
| Date of Award: | 2026 |
| Depositing User: | Theses Team |
| Unique ID: | glathesis:2026-86336 |
| Copyright: | Copyright of this thesis is held by the author. |
| Date Deposited: | 11 Nov 2026 12:09 |
| Last Modified: | 11 Nov 2026 12:24 |
| Thesis DOI: | 10.5525/gla.thesis.86336 |
| URI: | https://theses.gla.ac.uk/id/eprint/86336 |
| Related URLs: |
Actions (login required)
![]() |
View Item |
Downloads
Downloads per month over past year

Tools
Tools