Modelling motor learning in the elderly using a serious virtual reality game

Details

Supervisors

Massart, Estelle ; Everard, Gauthier

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

enThis work aims to study the evolution of subjects' performance in an experiment in immersive virtual reality relying on a serious game, inspired by the Whac-a-Mole arcade game. This was performed by modelling two outcomes: the action time required to hit the target and the probability of a successful hit. To account for repeated observations within subjects, mixed-effects models with a random intercept per subject were used. Separate models were fitted for the first and second halves of the experiment in order to compare early and late performance. The action time results showed that subjects tended to become faster as the experiment progressed, when task difficulty and subject baseline differences were taken into account. The late model achieved lower prediction errors than the early model in the repeated 80--20 test procedure, suggesting that action times became more predictable in the second half of the experiment. For the binary hit outcome, logistic mixed-effects models were used to predict the probability of hitting the target. The models achieved moderate predictive performance: they ranked successful trials above unsuccessful trials better than chance, but trial-level success remained difficult to predict precisely. The early model performed slightly better than the late model at the trial level, whereas the block-level bias remained close to zero in both cases. Per-subject analyses showed that some subjects were well predicted, while others had abrupt drops in hit rate or irregular performance patterns that were not fully explained by the available predictors. Finally, two model exploitation and generalisation procedures were performed: leave-one-subject-out analyses and difficulty-level simulations. The action time models generalised better than the binary hit models to unseen subjects. The simulations suggested that the late models generally predicted shorter action times for some subjects, and higher hit probabilities for all of them under comparable task configurations, but the magnitude of these changes varied between subjects.