Schedule of Events | Search Abstracts | Symposia | Invited Symposia | Poster Sessions | Data Blitz Sessions

Poster F81

Machine learning based prediction of mental fatigue based on eye-tracking and psychological data

Poster Session F - Tuesday, April 16, 2024, 8:00 – 10:00 am EDT, Sheraton Hall ABC

András Matuz1 (, Gergő Jakóczi1, Rebeka Gőgös1, András Zsidó1, Árpád Csathó1; 1University of Pécs

Prolonged performance of cognitively demanding tasks often leads to a psychobiological state labelled as mental fatigue. Fatigue has been associated with altered biological processes, for example, changes in cardiac activity or pupil size. Using eye-tracking and psychological data, we trained classification and regression algorithms to differentiate between fatigued and non-fatigued states and to predict the level of post-task fatigue, respectively. In two experiments, fatigue was induced by prolonged cognitive tasks. The difference between the two experiments was that cognitive load was higher in Experiment 1 (n = 30) than in Experiment 2 (n = 29). For classification, eye-tracking and psychophysical variables were calculated for the first (i.e. “non-fatigued” label) and the last 5-min of the task (i.e. “fatigued” label). For regression, demographic, eye-tracking, psychophysical and psychological variables were used to predict post-task fatigue. The best performing classification algorithm was the k-nearest neighbour algorithm (AUC = 0.612 [CI95% = 0.588–0.635]). Elastic net regression showed the best predictive performance among the regression algorithms (R2 = 0.608 [CI95% = 0.578–0.639]). Based on permutation importance, the best predictors of post-task fatigue were pupil size, fixation instability and sleep duration. To conclude, the regression-based prediction of post-task fatigue was successful and suggested that eye-tracking metrics explain individual differences in fatigue caused by prolonged cognitive performance. Classification algorithms, however, performed just slightly above chance level, which is a poorer performance compared to models trained on other biomarkers (e.g. heart rate). Support: AM was supported by the National Research, Development and Innovation Office, NKFIH grant (PD147001)

Topic Area: METHODS: Electrophysiology


CNS Account Login


April 13–16  |  2024