A neuroergonomic approach to performance estimation in a psychomotor vigilance task
Résumé
Introduction: Passive brain-computer interfaces (pBCI; tools that enable an implicit mental state
estimation) have gained attention in a wide range of applications, including performance and vigilance
monitoring in high-risk work settings (Lotte & Roy, 2019). Vigilance can be defined as the ability to
maintain sustained attention to a stimulus for an extended period of time (Al-Shargie et al., 2019), and is
influenced by the time of day and fatigue (Lim and Dinges, 2008). A vigilance decrement impacts
performance over time (called time-on-task -TOT) during tedious monitoring tasks, resulting in slower
reaction times or increased errors (Pattyn et al., 2008). This effect is experienced in all kinds of activities
such as in aeronautics where pilots can experience a performance drop during the flight (Wiggins, 2011).
Hence, vigilance and performance estimation is a crucial step towards the implementation of safer work
settings. Machine learning applied to physiological measures, such as cerebral activity (via
electroencephalogram -EEG- recordings), is a promising way to estimate performance. EEG’s spectral
activity is impacted by fluctuations in vigilance (Matousek & Petersén, 1983) and the power in both theta
and alpha bands can be considered as robust biomarkers of mental fatigue (Tran et al., 2020). Numerous
studies have attempted to estimate performance during a vigilance task based on EEG measures (Tian et
al. 2018), or electrocardiographic (ECG) measures (e.g., heart rate -HR-, and its variability -HRV; Chua et
al., 2012). To our knowledge, performance estimation during monotonous tasks has not reached a high
accuracy, and pBCI pipelines could be improved. Hence, the objective of this study is to employ a
comprehensive neuroergonomic approach to vigilance and performance characterization for a typical
vigilance task: the Psychomotor Vigilance Task (PVT; Dinges & Powell, 1985) encompassing statistical
analyses, as well as EEG-based performance classification using pre-stimulus signal.
Methods: Ten volunteers (3 females; Mage=25, sd=3; ethical number from Univ. Toulouse: 2021-342)
performed a 10-minute PVT (i.e., 90 stimuli) from a task battery. They had to complete a fatigue
questionnaire (Karolinska Sleepiness Scale; Åkerstedt & Gillberg, 1990) and their response time (RT) was
measured. In addition, EEG and ECG activities were recorded using an ActiCHamp system (63+1
electrodes). Data were processed in two ways: i) TOT analysis: data were split into ten 1-minute
windows; and ii) Performance analysis: response-based epoching (-2:0s for EEG; -10:0s for ECG), only the
30 best and 30 worst trials were kept (labeled according to RT), and also used for performance
classification. Considering statistical analyses, one-way repeated measures ANOVAs and paired samples
t-tests (Student and Wilcoxon), were performed separately on each dataset. To perform TOT analysis on
RT, the signal was cut into 10 periods of 9 simultaneous stimuli and the reciprocal RT (mean 1/RT) was
calculated on each period. EEG data were filtered (1-40 Hz) and ocular artifacts were automatically
removed (SOBI method). The theta, alpha, and beta power were extracted for three electrode clusters
(frontal, central, posterior). The Task Load Index (TLI: thetaFz/alphaPz) and the engagement ratio
(beta/[theta+alpha]; average on all electrodes) were also calculated. ECG data were filtered (1-40 Hz)
and normalized using a 1-minute eyes-open resting state period. HR and HRV (as SDNN) were then
extracted. For EEG-based estimation, a dimensionality reduction method based on Laplacian was applied
(Xu et al., 2021), and a minimum distance to mean with geodesic filtering classifier (FgMDM) was trained
and tested using a 10-fold cross-validation procedure.
Results: There was no significant difference in subjective fatigue (regardless of PVT order in battery).
There was a significant linear downward trend on reciprocal RT (p<.05; other contrasts n.s.; Fig.1.A).
There was a significant effect of TOT on alpha power at frontal (p<.05, ηp²=.26) and posterior (p<.05,
ηp²=.27) sites, as well as on the engagement ratio (p<.05, ηp²=.31; Fig.1.B). TOT also significantly impacted
HR (p<.001, ηp²=.31; Fig.1.C) and HRV (p<.05, ηp²=.27). Regarding performance (best/worst trials), alpha
power was significantly lower for the best than worst trials at frontal (p<.01, rB=-.93), and posterior
(p<.01, rB=-.89) sites, while the TLI was higher for the best trials (p<.05, d=.87). In terms of performance
classification, the FgMDM classifier achieved a mean accuracy of 58.2% without dimensionality
reduction. By selecting the number of dimensions that gives the best accuracy for each subject, the
mean accuracy with dimensionality reduction was 70.5% (mean number of dimensions = 27.3, Fig.2.A).
By comparison, the principal component analysis (PCA) achieved a mean accuracy of 71.1% (mean
number of dimensions = 23).
Discussion: The results showed that the neuroergonomic approach employed in this study enabled us to
assess physiological modulations due to vigilance fluctuations, such as alpha power and engagement
ratio decreased over time. However, the HR drop observed in the first minutes may be due to the initial
presentation novelty (Kelsey et al., 1999). Also, analyses of the best and worst trials showed that
vigilance decreased during the bad trials. However, some results were not as strong as those obtained in
the literature. This may be due to either the short task duration (10 minutes) compared to the several
sessions performed over several hours or days in the literature, or because our participants were not
sleep deprived. Besides, these analyses were performed on a small number of participants (n=10).
Regarding performance estimation, the classification accuracy was low for several reasons. Firstly, there
is little difference in terms of vigilance during the 10-minute PVT task, even though we have chosen the
30 best and worst trials. Secondly, the number of training samples is quite limited (i.e., 60 per subject). In
this case, dimensionality reduction becomes essential as shown by the results. Both dimensionality
reduction methods (non-supervised) significantly improve accuracy while reducing the computation
time. Moreover, on average the Laplacian-based method performs better in lower dimensions than PCA
(Fig.2.B). In summary, this study shows that even with a short task and a normal level of fatigue, it is
possible to observe an impact of a monotonous task on behavioral and physiological measures at the
group level. Yet, it remains difficult to implement an accurate performance estimation pipeline using a
single short session at the individual level.
Domaines
Neurosciences
Origine : Fichiers produits par l'(les) auteur(s)