*1.1. Background*

#### 1.1.1. Physiological Responses

The most common approach in stress sensing is interpreting physiological responses, such as EDA [29], heart rate [30] (HR), HRV [31], pupil dilation [32] (PD), skin temperature [33,34] (ST) or EEG [6,8,35]. In addition, muscle activity, such as microvibrations [9] and muscle tension [10,36], is affected by stress. The Sympathetic Nervous System (SNS) of the Autonomous Nervous System unconsciously controls these vital signs and are thus considered reliable sources [37]. Prior work demonstrated EDA as being linearly related to arousal and widely used in the context of stress sensing [38–40]. More specifically, EDA has been used to measure stress in applications, such as measuring the stress of call centre agents [39] and the discrimination of stress from the cognitive load [40]. In addition, in many laboratory and field studies, EDA is considered one of the gold-standard methods to sense stress [41]. As the SNS mainly controls EDA, it is regarded as a reliable physiological sign for acute stress [42].

Furthermore, prior work revealed that mental stress related to cognitive load has an impact on HRV [6,8], in particular towards reduced HF components [43–45]. Measurable changes in HR has also been observed during high attention tasks [44]. Electrocardiography (ECG) and Photoplethysmogram (PPG) [46,47] are the earliest technologies used in literature to measure HR and HRV. High power consumption and tight sensor placements are the major draw back of these technologies. Meanwhile, various studies, such as BioWatch [48], SenseGlass [49], SeismoTracker [50], investigated smart devices capable of measuring HR and HRV based on ballistorcardiography (BCG). Although BCG can compete with state-of-the-art techniques [51], it is susceptible to unwanted motion artifacts and thus only reliable in resting states, such as sleeping, standing or sitting. In summary, wearable sensing using wristbands, nail clips, vests and headbands are often uncomfortable, given the bulkiness and tight sensor mounting.

Another option is contact-free sensing of physiological data [52], such as using thermal cameras [53,54] or Doppler radar [55]. As cameras are typically expensive, webcams are a low-cost alternative to measuring HR and HRV from the human face [17,18]. Inspired by previous studies, COGCAM [16] measured cognitive stress with digital cameras, placed 3m away from the user. In these studies, participants are instructed not to move their head, which restricts natural behaviours. An ambient light source ensuring constant light conditions is essential. Privacy concerns may arise when using cameras for tracking.

An increased muscle activity can also indicate stress. For instance, an increased amplitude of the muscles' microvibrations [9] and an increased muscle tension can provide stress indicators [10]. Other researchers [41,56,57] have detected some types of muscle tension implicitly by analysing keyboard and mouse control. An increased typing pressure and a greater contact with the surface of the mouse has been observed in stressed conditions [41]. We believe such implicit sensing is an unobtrusive and thus desirable approach we would also like to explore.

#### 1.1.2. Facial Expressions

In recent years, identifying affective states based on facial expressions has been widely explored in the area of affective computing [58]. Technology-wise, vision-based camera tracking, including depth cameras [59], are the most commonly used technologies in identifying facial expressions [60]. A recent work [61] investigated stress and anxiety sensing using facial cues, in which the authors' induced acute stress by internal and external stressors. All videos recorded were analysed posterior and a machine learning model scored between 80–90% accuracy in recognizing stress tasks. Although vision-based identification is demonstrably effective, unfavourable light conditions and movement artifacts easily affect detection accuracy.

Other researchers identify facial expressions based on skin-contact electrodes, such as piezoelectric sensing [62], EMG [63,64], capacitive sensing [65,66] and electric field sensing [67]. These researches demonstrate the detection of facial gestures, such as frowning, eye wink, eye-down, mouth movements, etc. and infer on emotions such as frustration, confusion and interest engagement. Although the identification of emotions and facial expressions seem reliable, the outcome may differ in real-life scenarios given that facial expressions are known to deceive easily [68].
