*6.2. Applications*

#### 6.2.1. Professional Environment

In our studies, the recruited participants mainly performed computer-aided tasks in an office environment while in a sitting posture. Other occupations, such as cashiers, emergency/non emergency call centre workers, also perform their tasks while in sitting posture for prolonged periods. In addition, the working environment of these occupations are similar to an office environment. Hence, our method may work for these occupations as well. In such a scenario, stress sensing could be used to improve mental health. However, further studies need to identify the viability of foot-based stress sensing method for such occupations. In addition, remote stress monitoring of adults under home care is a another potential application. Simple accelerometer can be embedded to a sock and thus monitor both stress and activity level using a single sensor. However, to accomplish this further studies with better classification algorithms stated in literature are required [98,99].

#### 6.2.2. The Quantified Self

Enriching self-tracking with a stress detection is another application. In this community, stress is being triangulated with several data of wearable sensors [100]. A simple feature could substantially increase accuracy. Moreover, using an already available wearable, like a shoe, can address users who prefer not to wear additional garmen<sup>t</sup> accessories.

#### *6.3. Limitations and Future Work*

#### 6.3.1. Quantifying Multiple Stress Levels

Both lab studies only investigated discriminating stress from relaxation without aiming to identify different levels of stress. However, the results of study 3 indicates that the frequency of demonstrating stress related postures could reveal the extent of stress. However, it requires further research to infer on different levels of stress that could be based on the frequency of stress-related foot movements and foot posture characteristics.

#### 6.3.2. Stress Detection in Sitting Posture and Other Activities

In our society, sitting is the most common posture demonstrated during both the working week and weekends [101], which our findings also confirm (see Table 3). Therefore, our proposed models would principally work for a majority of the time, whenever an individual is doing some sort of intellectual work while seated. Detecting stress in other postures exceeds the scope of our current research. Identifying stressful situations when performing regular activities, such as standing, walking and running are thus considered future work. For this, we would first need to identify the current posture and activity, such as by using an insole [24,25] or IMU [92,94,102], as shown in the literature.

#### 6.3.3. Accuracy Boost with Personalised Models

We aimed to develop a generalised model capable of working across different users. We observed that the accuracy of our generalised models decreases for features, such as Median Frequency and Dominant Frequency. This is because not all users demonstrate an elevated foot tapping/shaking when stressed. In addition, we observed that individuals may have slightly different foot posture and motion characteristics depending on the BMI and other personal habits. Relying on a personalised model will further boost accuracy, particularly for features like foot-tapping. Further, a Neural Network approach can be advantageous, particularly when building a personalised model with a high volume of data gathered in the wild.

## 6.3.4. Improved Hardware

In our current prototype, the IMU (in conjunction with the control unit) was attached to the ankle. Our long-term goal is to integrate all these components into a smart footwear. This may change the orientation of the IMU. Thus, the main signal could spread to different axis, resulting in another axis to provide higher separation sharpness.
