Frontal EEG-Based Multi-Level Attention States Recognition Using Dynamical Complexity and Extreme Gradient Boosting

Abstract

Measuring and identifying the specific level of sustained attention during continuous tasks is essential in many applications, especially for avoiding the terrible consequences caused by reduced attention of people with special tasks. To this end, we recorded EEG signals from 42 subjects during the performance of a sustained attention task and obtained resting state and three levels of attentional states using the calibrated response time. EEG-based dynamical complexity features and Extreme Gradient Boosting (XGBoost) classifier were proposed as the classification model, Complexity-XGBoost, to distinguish multi-level attention states with improved accuracy. The maximum average accuracy of Complexity-XGBoost were 81.39 ± 1.47% for four attention levels, 80.42 ± 0.84% for three attention levels, and 95.36 ± 2.31% for two attention levels in 5-fold cross-validation. The proposed method is compared with other models of traditional EEG features and different classification algorithms, the results confirmed the effectiveness of the proposed method. We also found that the frontal EEG dynamical complexity measures were related to the changing process of response during sustained attention task. The proposed dynamical complexity approach could be helpful to recognize attention status during important tasks to improve safety and efficiency, and be useful for further brain-computer interaction research in clinical research or daily practice, such as the cognitive assessment or neural feedback treatment of individuals with attention deficit hyperactivity disorders, Alzheimer’s disease, and other diseases which affect the sustained attention function.

Signal Quality Investigation of a New Wearable Frontal Lobe EEG Device

“…Although attention recognition is not the main purpose of this study, the accuracies of distinguishing between high and low attention in this study are 71.12% and 74.76%, which is lower than previous studies [37]. The reasons may be: (1) the duration of AX-CPT was only about 6 min, and the subjects were healthy college students, and their attention may always be kept at a high level during the AX-CPT task; (2) only the four prefrontal lobe channels were collected in this study, and the number of channels was small; and…”

“…The four-channel average relative PSD of Mindeep in the whole frequency band (0.3-45 Hz) was similar to Neuroscan in EC and EO. It can be observed that the EO state had higher delta (0.3-4 Hz) and alpha (8-13 Hz) relative PSD, and lower beta (13-30 Hz) and gamma (30)(31)(32)(33)(34)(35)(36)(37)(38)(39)(40)(41)(42)(43)(44)(45) relative PSD. The relative power of delta, theta, alpha, beta, and gamma bands were shown in Figure 6b-d.…”

“…All participants were required to sit on the chair with their eyes about 0.5 m away from the screen. The experiment included three parts (see Figure 3b-d): (1) resting task; (2) auditory oddball task [39]; and (3) attention task: AX continuous performance test (AX-CPT) [37,40].…”

“…where, B m (r) is the number of template vector pairs having d[X m (i), X m (j)] < r and represents the similarity between two sequences of length m, A m+1 (r) is the number of template vector pairs having d[X m+1 (i), X m+1 (j)] < r and represents the similarity between two sequences of length m + 1. According to our previous studies [31,37], m = 2 and r = 0.2 were used in this study.…”

“…The SVM model performed the best, with an average accuracy rate of 91.72%. In our previous study [37], attention was divided into three levels: high, medium, and low, according to performance on the sustained attention task. Comparing the three classification results of SVM, RF, and XGBoost model, the XGBoost model showed the best performance with an accuracy of 80.42%.…”