Abstract

Mental stress is considered as an essential element that affects decision making. Apart from mental stress, cognitive workload, mental effort, attention, and cognitive engagement are also involved in the decision-making process. Ambiguities of these concepts lead to confusion in their applications.
One objective of this thesis is to explore the relationship between mental stress and stress-related concepts. By investigating the mechanisms for decision-making, the difference and correlation of mental stress and other concepts are disclosed.
Heart rate variability (HRV) is a common method to measure mental stress. By investigating the correlation between HRV and mental stress, it can be confirmed that HRV does respond to mental stress changes instead of other concepts. HRV features are used to assess whether there is a relationship between baseline HRV and mental stress. However, the extracted features usually contain a large amount of redundancy, which adds computational complexity to mental stress quantification while not contributing to quantification accuracy. Recently, researchers have resorted to the random forest as a tool for HRV feature selection.
Another objective of this thesis is to select significant HRV features to quantify the mental stresses using the random forest method.
In this thesis, an open-source data set, called the SWELL-KW data set, is used for mental stress measurement, where three labels are assigned according to different mental stress conditions, i.e., neutral, time pressure, and interruption. A set of HRV features are proposed based on time domain and frequency domain analysis for mental stress measurement. Statistical analysis is performed to select the essential features that reflect mental stress.
The random forest algorithm of feature selection is then studied, and the accuracy in measuring mental stress is validated by comparing the extracted features of the training set and the testing set. In order to evaluate the random forest algorithm's performance, the comparisons with other related algorithms, including support vector machine (SVM), decision tree, gradient boosting decision tree (GBDT), k-nearest neighbor algorithm (KNN), and deep neural networks (DNN), are also conducted in terms of accuracy and time cost.
The optimal HRV feature subset is proposed for mental stress quantification, including median RR, mean RR, median REL RR, HR, pNN25, SDRR RMSSD, SDRR RMSSD REL RR, TP, SD2, and SDRR. It is shown that this subset of features gives a high feature importance score and thus has a significant effect on mental stress quantification.
Performing random forest analysis with a sufficient amount of labeled data shows that the optimal HRV feature subset yields high mental stress quantification accuracy by using random forest. Moreover, random forest always makes the best overall performance in feature selection compared with other algorithms in terms of accuracy and time cost. It also infers the potential relation between physiological responses and mental activities.