Strong evidence from research suggests that physiological signals carry information about human emotions [16]. A fast heartbeat, excessive sweating, and unusual facial expressions are typical manifestations of emotion, which is an intense mental experience [17]. According to the results shown in [18], a lower HRV is linked to the feeling of “happiness”, whereas a higher value is linked to “joy” or “amusement.” In a similar vein, research has also been done on the function of GSR. It is shown that GSR can distinguish between the emotions of “fear” and “anger”, “fear” and “sadness”, and “happy” and “sad” [19].

Mental stress has long been recognized to have harmful effects on human health. Continuous mental stress, for instance, can lead to physical and mental conditions such as cardiovascular diseases, hypertension, diabetes, cancer, headaches, depression, anxiety, and insomnia. Early detection of high stress levels is necessary to stop these harmful effects. Several recent stress-detection strategies have been put forth, typically based on machine-learning (ML) methods. The work of Kyriakou et al. [21] in 2019 aimed to bridge the gap between laboratory settings and real-world field studies by introducing a new algorithm to detect moments of stress (MOS) using wearable physiological sensors. Eleven subjects wore an Empatica E4 device and were subjected to a laboratory experiment, an auditory stimulus was used to induce stress. Furthermore, to validate the algorithm, a real-world urban experiment was introduced. An accuracy of 84% was obtained using the proposed algorithm. The study carried out by Kaczor et al. [22] aimed at the objective measurement of physician stress in the emergency department using the Empatica E4 smartwatch. EDA, acceleration, and heart rate signals were acquired from eight participants during clinical shifts (typically 8–10 h). After that several machine-learning classifiers were used and the best accuracy obtained was 70% to detect stress during the working shift with respect to the baseline condition. In the study proposed by Dai et al. [23], 32 subjects participated in 2 h of laboratory and 24 h of field-based experiments. The aim of the study was comparing between objective and subjective stress-detection models. In particular, participants were given a mental mathematics assignment, which required them to solve a series of complex mathematical problems over a given period of time. Support Vector Machine, Random Forest, AdaBoost, Gradient Boosting, and Logistic Regression classifiers were used to detect stressed or non-stressed periods in both objective and subjective stress models. In the study carried out by Mach et al. [24] in 2022, a laboratory experiment consisting of an arithmetic task which is counting down or up steadily, and physical activity (sitting vs. stepping) with 52 participants was conducted. This study aimed to assess mental workload via heart rate measurement and a chest strap with a 1-channel ECG. They found that the mean heart rate increased when participants performed the arithmetic task compared to the conditions with no arithmetic task while sitting and stepping. In the study done by Seo et al. [25] in 2022, 24 participants wore a Zephyr chest strap equipped with a BioHarness module to acquire ECG and Respiratory signals. Furthermore, the subjects were sitting in front of a laptop and faced a camcorder screen to register facial information. The experiment lasted for 45 min and comprised two stages: an initial setting stage, and an actual experiment stage which is the Stroop task. The actual experiment consists of Relax, Easy Stroop, Recovery, Hard Stroop, and Recovery, 5 min for each. Afterwards, signal and image processing was done followed by a Deep Neural Network (DNN) classifier. The accuracy for two or three levels of stress classification was 73.3%, and 54.4%, respectively.

In a study by Chalabianloo et al. [26] in 2022, 32 subjects were subjected to a laboratory experiment that consisted of baseline, stress, recovery, and cycling sessions. Stress sessions were performed using the Stroop task and different physiological signals were recorded using seven different wearable devices simultaneously. The best accuracies across most of the devices were obtained using an Extremely Randomized Tree classifier, for example, 88.26% for the BITalino device. Furthermore, to study the effects of multimodality, the EDA signal was introduced using Empatica E4. After that, the same classifiers mentioned above were used. The accuracy obtained considering only HR was 83.89% using the Random Forest classifier, while when considering HR and EDA the accuracy became 90% using the Extremely Randomized Tree classifier.

In the study done by Suni Lopez et al. [27] a laboratory experiment was conducted to detect stress in the office workplace, the experiment consisted of interacting with a laptop where the Stroop task was installed. Twelve subjects participated and were asked to wear the E4 smartwatch to collect EDA data, and headphones to interact with the environmental trigger (fire alarm). After signal filtering, aggregation, and discretization, an accuracy of 79.17% was obtained using statistical method classification.

In Table 1, all the mentioned works together with the devices and protocol applied are summarized.