Selecting Feature Sets and Comparing Classification Methods for Cognitive State Estimation

Ultra-Short Window Length and Feature Importance Analysis for Cognitive Load Detection from Wearable Sensors

“…Because people do not recover instantaneously, physiological reactions caused by cognitive tasks are still ongoing when the rest session begins, which likely affected the classification performance. A significantly higher classification performance was found, e.g., in [20], where the resting condition represented a baseline measurement conducted at the beginning and the end of the measurement protocol. However, these kinds of baseline rest periods have not been recorded in this dataset.…”

“…The features were normalized using within-subject standardization, meaning that each feature was transformed by subtracting its mean and dividing by its standard deviation separately for each participant. Person-specific standardization was conducted instead of person-independent standardization, since it has shown improved performance in earlier work in similar contexts [14,20,37].…”

“…Because the number of samples in each class is reasonably balanced (approximately 46% rests and 54% cognitive loads for each window length), the classification performance was assessed in terms of accuracy, the percentage of correctly classified samples. Extreme Gradient Boosting [38] (XGB) was selected as the classification model for the following reasons: the classifiers from the random forest family and the boosting method have been shown to be strong classifiers [39], XGB has shown good performance earlier in similar contexts [14,20] and because XGB is computationally efficient and scales to very large datasets [38]. XGB is an ensemble of decision trees, each of which splits the data hierarchically, aiming to contain data originating from a single class in each leaf node.…”

“…The changes in the cognitive load are also reflected in various biosignals that can be measured by using biosensors, e.g., wearable devices [13,14]. For instance, increasing task difficulty (or cognitive load) and acute stress increases the HR and breathing rate [15], ST [11] as well as GSR [16] and decreases the HRV [17], number of eye movements [18], and increases the blink rate [19,20].…”

“…In addition, there are few studies where the effect of window length to classification accuracy has been studied (see Table 1) and even those have mainly used windows with length of 30 s or more. [25,27,29], and 1000 Hz [9,20]. Abbreviations: ACC: acceleration, (d)BN: (dynamic) Bayesian network, DT: decision tree, EEG: electroencephalogram, EOG: electro-oculogram, GSR: galvanic skin response, HR: heart rate, HRV: heart rate variability, kNN: k-nearest neighbors, LDA: linear discriminant analysis, NA: not available, QDA: quadratic discriminant analysis, RESP: respiration, RF: random forest, SOM: self-organizing map, ST: skin temperature, SVM: support vector machine, XGB: extreme gradient boosting.…”