*5.1. Statistical Data Analysis*

### 5.1.1. Validation of Different Perceived Stress Levels by using the Self-Reports

In order to validate that the participants experienced different perceived stress levels in different contexts (lecture, relaxation, presentation), we used the Frustration item (see Section 4.5) from the NASA-TLX [57]. The distribution of answers is demonstrated in Figure 7. Our aim is to show that the perceived stress levels (obtained from self-report answers) differ in relaxation sessions considerably when compared to the presentation session (high stress). To this end, we applied the *t*-test (in R programming language) to the perceived stress self-report answers of yoga versus presentation, mindfulness versus presentation and pause (mobile mindfulness) versus presentation session pairs. The paired *t*-test is used to evaluate the separability of each session. The degree of freedom is 15. We applied the variance test to each session tuple; we could not identify equal variance in any of the session tuples. Thus, we selected the variance as unequal. We used 99.5% confidence intervals. The *t*-test results' (*p*-values and test statistics) are provided in Table 5. For all tuples, the null hypothesis stating that the perceived stress of the relaxation method is not less than the presentation session is rejected. The perceived stress levels of participants for all meditation sessions are observed to be significantly lower than the presentation session (high stress).

**Table 5.** *T*-test results for session tuple comparison of perceived stress levels using self-reports.


**Figure 7.** Visual representation of the frustration scores collected in different types of sessions.

5.1.2. Before and After Physiological Measurements for Evaluating Performance of Yoga and Mindfulness with Blood Pressure

In this section, we compared the effect of stress managemen<sup>t</sup> tools such as yoga and mindfulness on blood pressure. It is expected that blood pressure sensors will be part of unobtrusive wrist-worn wearable sensors soon. We plan to integrate a blood pressure (BP) module to our system when they are available. Therefore, by using the measurements of a medical-grade blood pressure monitor, we provided insights about how stress reaction affects BP. We further applied and tested the prominent emotion regulation model of James Gross by analyzing these measurements in the context of stress management. We measured the diastolic and systolic BP and pulse using a medical-grade blood pressure monitor before and after the yoga and mindfulness sessions. In order to ensure that the participants were relaxed and that an accurate BP was recorded, BP was measured three times with the mean as the recorded result. A one-sample *t*-test was applied to the difference between mean values. The results are shown in Table 6.

Mindfulness decreased the systolic BP, –1.13% (ns), increased diastolic BP, +1.75% (*p* < 0.05) and decreased the pulse –5.75% (*p* < 0.05). Medicine knows that systolic blood pressure (the top number or highest blood pressure when the heart is squeezing and pushing the blood around the body) is more important than diastolic blood pressure (the bottom number or lowest blood pressure between heartbeats) because it gives the best idea of the risk of having a stroke or heart attack. In this view, the significant reduction of systolic BP after mindfulness is an important result.

Moreover, the difference between systolic and diastolic BP is called pulse pressure. For example, 120 systolic minus 60 diastolic equals a pulse pressure of 60. It is also known that a pulse pressure greater than 60 can be a predictor of heart attacks or other cardiovascular diseases, while a low pulse pressure (less than 40) may indicate poor heart function. In our study, pulse pressure was lower after mindfulness (we had both a significant reduction in systolic BP and an increase in diastolic BP), but its value was higher than 40 (42.69 mean difference before the mindfulness and 40.48 mean difference after the mindfulness), suggesting that this result can also be considered clinically positive.

During yoga, there was a decrease in systolic BP by −5.81% (*p* < 0.05), diastolic BP by −1.93% (ns) and increase in pulse +8.06% (*p* < 0.05). Yoga appears to be more effective than mindfulness at decreasing systolic and diastolic blood pressure, although mindfulness seems to be more effective than yoga for decreasing the pulse due to the activity involved in yoga.

**Table 6.** The difference between the mean diastolic blood pressure, the mean systolic blood pressure and the mean pulse, before and after sessions of guided mindfulness and guided yoga. (\* *p* < 0.05).


### *5.2. Physiological Stress Level Detection with Wearables by Using Context Labels as the Class Label*

We tested our system by using the known context labels of sessions as the class label. We used Lecture (mild stress), Yoga and Mindfulness (relax) and Presentation in front of the board of juries (high stress) as class labels by examining perceived stress self-report answers in Figure 6. We investigated the success of relaxation methods, different modalities and finding the presenter.

### 5.2.1. Effect of Different Physiological Signals on Stress Detection

We evaluated the effect of using the interbeat-interval, the skin conductance and the accelerometer signals separately and in a combined manner on two and three class classification performance. These classes are mild stress, high stress and relax states from mindfulness and yoga sessions. The results are shown in Tables 7–9. For the three-class classification problem, we achieved a maximum accuracy of 72% by using MLP on only HRV features and 86.61% with only accelerometer features using the Random Forest classifier and 85.36% accuracy combination of all features with LDA classifier (see Table 7). The difficulty in this classification task is a similar physiological reaction to relax and mild stress situations. However, since the main focus of our study is to discriminate high stress from other classes to offer relaxation techniques in this state, it did not affect our system performance. We also investigated high-mild stress and high stress-relax 2-class classification performance. For the discrimination of high and mild stress, HRV outperformed other signals with 98% accuracy using MLP (see Table 8). In the high stress-relax 2-class problem, only HRV features with RF achieved a maximum accuracy of 86%, whereas ACC features with MLP achieved a maximum of 94% accuracy. In this problem, the combination of all signals with RF achieved 92% accuracy which is the best among all classifiers (see Table 9). For all models, EDA did not perform well. This might be caused by the loose contact with EDA electrodes in the strap due to loosely worn smartbands.


**Table 7.** Effect of different modalities and their combination on the system performance. Note that the number of classes is fixed at 3 (high stress, mild stress and relax).


**Table 8.** Effect of different modalities and their combination on the system performance. Note that the number of classes is fixed at 2 (high stress and mild stress).

**Table 9.** Effect of different modalities and their combination on the system performance. Note that the number of classes is fixed at 2 (high stress and relax).


5.2.2. Effectiveness of Yoga, Mindfulness and Mobile Mindfulness (Pause)

We applied three different relaxation methods to manage stress levels of individuals. In order to measure the effectiveness of each method, we examined how easily these physiological signals in the relaxation sessions can be separated from high stress presentations. If it can be separated from high stress levels with higher classification performance, it could be inferred that they are more successful at reducing stress. As seen in Tables 10 and 11, mobile mindfulness has lower success in reducing stress levels. Yoga has the highest classification performance with both HR and EDA signals.

**Table 10.** The classification accuracy of the relaxation sessions using stress managemen<sup>t</sup> methods and stressful sessions using EDA.



**Table 11.** The classification accuracy of the relaxation sessions using stress managemen<sup>t</sup> methods and stressful sessions using HRV.
