Skip to Main content Skip to Navigation

Détection de la fatigue mentale à partir de données électrophysiologiques

Abstract : Three experiments were conducted to assess the feasibility of a mental fatigue detector that would exploit electrophysiological signals for discriminating time windows associated with a "fatigue" state of windows associated with a "no fatigue" state. In the first experiment, subjects performed a task switching task for about 2 h. Two segments of 20 min were selected for each subject. The first one was associated with the "no fatigue" state and the other one with the "fatigue" state. These segments were divided into sliding time windows and electrophysiological signals (electro-oculo-, cardio-, encephalograms - EOG, ECG and EEG) were quantified within each time window. Using linear support vector machines (SVMs - trained on a subjet-by-subjet basis) to classify the quantified time windows allowed several observations. Correct classification rates increased with time-window length (4 to 30 s long) but not information transfer rates (ITRs). ECG quantifications performed better than EOG quantifications. EEG mean amplitudes yielded higher classification scores than EOG or ECG quantifications, and EEG phase locking values (PLVs) in turn outperform EEG mean amplitudes with classification rates higher than 90 %. Review of SVM weights showed that EOG and merely ECG quantifications contributed to the classification where all the quantifications of the different modalities were used together. EEG mean amplitudes increased in the 7-13 Hz frequency range and decreased in the 13-18 Hz band. 3-7 Hz mean amplitudes increased in FCz, evoking the error response negativity (ERN). In a second experiment involving a spatial compatibility task, we discriminated 20-s-long windows between higher (shorter mean reaction time) and lower performance levels. The signal in these windows was quantified using mean amplitude, PLV or coherence at electrodes and reconstructed sources in the cerebral cortex. We expected higher classification rates from reconstructed data than from scalp data, regardless of the quantification, but it was not the case. We also made a simulation to evaluate the approach above, and simulated two distinct phenomena: one involving a change in amplitude of two source signals, and another characterized by a change in phase coupling between the same two sources. These simulations showed that, on one hand, complementary of local synchrony measurements such as mean amplitudes, and distant synchrony measurements such as coherence or PLV. We also demonstrated how useful could be source reconstruction for correctly classifying and localizing discriminant patterns. This discrepancy between the results obtained from simulated data and real EEG data was interpreted as due to the contribution of extra-cortical sources that were not considered in existing forward models, such as subcortical sources and extra-cranial sources (eyes, muscles). The third study was based on a third experiment with an ecological context, and involved student pilots performing in a flight simulator. The simulation was about 3-h long and EEG signals together with 3 flight parameters (heading, flight level and airspeed) were recorded. We presented a new behavioral measure quantifying the variability of flight parameters over 10 s, useful for monitoring pilot performance with a temporal resolution suitable for our analyses. We designed two classes: a first group of 10-s time windows associated with higher performance (or lower variability) and a second group of windows associated with poorer performance. Instead of using a cross-validation procedure, we tested the trained classifiers on post-training data, in order to evaluate the ability of the classification procedure to identify a trend in pilot performance. This procedure indeed determined whether pilot performance deteriorated at the end of the simulated flight or not for 12 out of 13 pilots. The classification does not deteriorate with time between the end of the data used for training and the beginning of the data used to evaluate the classifier learned. Exhibited physiological correlates were not consistent from one experiment to another. Therefore, the different class designs may not have isolated a single mental fatigue differential, or mental fatigue may have depended of the task and/or subjects. However, the feasibility of discriminating between mental states in ecological contexts was demonstrated.
Document type :
Complete list of metadata

Cited literature [78 references]  Display  Hide  Download
Contributor : François Laurent <>
Submitted on : Sunday, October 23, 2011 - 1:50:54 PM
Last modification on : Wednesday, December 9, 2020 - 3:10:39 PM


  • HAL Id : tel-00634776, version 1


François Laurent. Détection de la fatigue mentale à partir de données électrophysiologiques. Neurosciences [q-bio.NC]. Université Pierre et Marie Curie - Paris VI, 2010. Français. ⟨tel-00634776⟩



Record views


Files downloads