3.1.1. Setup

The setup was done as follows;


3.1.2. Pre-Stress Measurements

> Before the experiment, the following procedure was applied;



**Figure 1.** The Perceived Stress Scale (PSS)-5 questionnaire used in the experiment.

## 3.1.3. The TSST

Our implementation of the TSST is described as follows:


be a good candidate for your ideal job. Your speech will be videotaped and reviewed by the psychologists that we conduct the research with. You have five minutes to prepare and your time begins now."


**Figure 2.** An example scene from the TSST phase in our experiment. The participant is presented at this moment in front of the neutral experimenter.

#### 3.1.4. Post-Stress Recovery Measurements

In order to alleviate the stress response, we applied a biofeedback based intervention, which was the built-in breathing application of Apple Watch [38]. The procedure was applied as follows:


#### *3.2. Daily Life Data Collection and Ecological Momentary Assessment*

After the controlled room experiments were finished, we gave the Empatica E4 devices to all participants. They were told to wear the Empatica E4 devices for twelve hours per day, between 9 a.m. and 9 p.m., for seven days. These days were not necessarily consecutive. We applied the EMA to collect information about the subject's stress level [39]. EMA involved the repeated sampling of subjects' current behaviors and experiences in real time, in subjects' natural environments [39]. EMA aims to minimize recall bias, maximize ecological validity, and allow the study of micro processes that influence behavior in real-world contexts. We implemented an online version of the PSS-5 questionnaire. For each three hour session, we asked the participant to fill in the EMA. In order to make sure the collection of self-reports, we sent them e-mails over seven days at the end of each three hour session when they were wearing the wristband. In other words, the participants were reminded to fill in the PSS-5 at 12 p.m., 3 p.m., 6 p.m., and 9 p.m. over seven days. The EMA was delivered to the participants through a survey app. The app was available on both desktop and mobile browsers. The link to the EMA was delivered to the participants through e-mail. Each e-mail that was sent in order to remind participants to fill in the EMA contained the link to the EMA. This questionnaire is strongly correlated with PSS-14 and appropriate for ambulatory settings. In total, we obtained 1003 h of physiological data and 388 EMAs (including 14 h of physiological data and 56 EMAs collected in the lab). There were some sessions with missing EMAs (60 EMAs in total), and we disregarded their physiological data.

The procedure of the methodology used in this study was approved by the Institutional Review Board for Research with Human Subjects of Bo ˘gaziçi University with Approval Number 2018/16. Prior to the data acquisition, each participant received a consent form that explained the experimental procedure and its benefits and implications to both society and the subject. The procedure was also explained verbally to the subject. All of the data were stored anonymously.

#### *3.3. Stress Recognition Framework*

In order to propose an unobtrusive stress detector suitable for everyday use, we used an Empatica E4 [40] comfortable wristband, which has more than 48 h of battery life and is equipped with a variety of sensors such as the three-dimensional accelerometer (32 Hz), the continuous heart rate monitoring unit based on the photoplethysmography (PPG), the skin temperature (4 Hz), and the EDA sensors (4 Hz). The major difference between the wristbands and contact sensors used in hospital settings is the vulnerability to the motion artifacts due to their design and attachment to the body. A daily life suitable and comfortable stress detection system should consider these artifacts. Thus, it should have a preprocessing unit that detects and removes the artifacts due to contact loss and motion. In daily life, the activity of the individual is important; for example, HRV and EDA can change due to high-intensity activity in short periods. Therefore a single sensor-based system can fail. Hence, a stress recognition system suitable for daily life should be multi-modal in terms of the collected data. We used robust preprocessing and feature extraction modules from our previous work [18] for this purpose (see Figure 3).

**Figure 3.** A high level block diagram of the stress level detection system with the Empatica E4 wristband.
