*6.1. Accuracy*

Our highest performing model is an LDA classifier with an accuracy of 86% in laboratory conditions. In literature, the models based on physiological parameters such as EDA, ECG, HRV, showed similar or higher accuracy in laboratory validations [32,39]. For example, an SVM model, based on features from a combination of physiological signals such as EDA, Blood Volume Pulse, ST and PD achieved an accuracy of 90.1% [32]. Although the model has achieved a higher accuracy, it has some limitations in field deployment. PD requires line of cite and may generate privacy concerns. In addition, according to authors, removing PD may drop the accuracy closer to 60%. In another study, authors were able to discriminate stress from cognitive load by using a LDA classifier based on EDA [40]. In a laboratory setting, they achieved an accuracy of 82.8%, which is slightly lower than the results we achieved in our study. In addition, an SVM model based on facial EMG, respiration, EDA and ECG was used to recognise 5 emotional states such as high stress, low stress, disappointment, euphoria and neutral. The paper reports an accuracy of 86% in a laboratory study [95]. In addition, another study reported a system which can classify stress with 86% accuracy based on 15 features extracted from EEG, ECG and EDA signal [96]. All these systems may be inconvenient, given the need to tightly attach multiple wearable sensors onto the body. In our approach, we can recognise stress and relaxation with a similar accuracy by using a simple accelerometer model, such as model C3. However, our current model cannot detect different levels of stress.

Some of the prior studies related to body language [76] study and facial [61] expression also achieved similar or slightly higher accuracy. However, many of these methods used camera-based systems, which may have limitations in real-life implementations. Some of the higher accuracy methods use multiple motion capture cameras, which is impractical to deploy in real-life, specifically in an office environment. For example, a work which detected emotions related to negative stress such as sadness, joy, anger and fear showed an accuracy of 93% [76]. However, the method uses 6 camera vicon motion capture system. Although the accuracy is slightly lower than vision-based methods, our approach captures certain body language related to stress while sitting by using a more practical method, which can be easily used in real-life applications.

On the other hand, there are several real life validations reported in literature. Healey and Picard proposed an LDA classifier to recognise stress of drivers using ECG, EMG, EDA and respiration [4]. Regardless of the cumbersome setup which consists of many on body sensors, they were able to recognise three levels of stress (low, medium and high) in 97% of accuracy. In another study, Hernandez et al. achieved an accuracy of 74% in call-center stress detection [39]. They proposed a person specific SVM model, which uses EDA response for stress detection. In our field study, by using four features related to foot motion and posture, we identified significantly high correlations between self-rated stress levels and model-derived stress levels across users, ultimately showing the robustness of the method.

Overall, methods which utilise a combination of physiological parameters seem to achieve higher accuracy than our proposed method [4,32]. However, sensing multiple physiological parameters requires attaching multiple sensors onto the user, compromising the comfort. Contrarily, previous methods based on single physiological parameter demonstrated either similar accuracy or lower accuracy [39,40]. The lower accuracy could be due to data losses and subjective differences. Some methods based on body language and facial expressions have shown higher accuracy [61,76] due to high sensing accuracy in visual-based sensing. However, those methods seem highly obtrusive in real-life.

In addition, a recent study compared unobtrusive sensors for stress detection at sedentary computer work [97]. In their analysis, they considered wrist worn, chest worn and thermal imaging based sensors. They identified that wrist worn sensors, such as EDA and PPG, may not capture stress accurately due to frequent data losses. This is mainly due to motion artefacts, such as electrode movements, detaching from the skin and a change in pressure on the skin. In addition, chest worn sensors which sense HR showed similar issues due to posture changes generating high noise and thus failing to maintain proper contact with the skin. This is highly problematic in 6–8 h of sensing in a typical working day. However, our approach does not result in such issues, specifically in an office working environment. We have proven that risks of data losses are not present with our method, given we sense foot motion and posture characteristics while sitting. While the study identified that thermal imaging resulted in a greater identification of stress during computer work, this is not always a practical method. Thermal imaging requires a consistent line of sight, which is problematic when the individual needs to attend to tasks away from their typical working desk. Our method poses no such issues.
