Electrodermal activity: A continuous monitor of well-being

1.Introduction

A smart society, society 5.0, envisions an iterative cycle which gathers data, analyzes it, and turns it into actionable insights, aiming for a human-centered society that balances economic progress and social problems, all while merging cyberspace with the physical world [18]. In other words, smartness is considered more than efficiency and effectiveness, and includes human understanding such as the ability to determine the psychological states of individuals [25]. A smart society can than adapt and support the individual if they require this. Through the understanding of psychological states, we can intervene accordingly and ultimately enhance individual well-being, creating flourishing societies [1]. Well-being is defined as a construct divided in two branches: hedonic (i.e., positive and negative affect) and eudaimonic (i.e., overall fulfillment) [10,29]. An example how smart society can help is signifying stressors and using positive cognitive reappraisal which can moderate the negative outcomes, improving hedonic state [27].

In this study, our particular interest lies in the assessment of prolonged stress levels experienced by individuals, which negatively influences well-being. One way in which this prolonged stress level can be expressed is through arousal (i.e. level of physiological activity) [1]. The comprehension of individual states can be achieved through diverse information channels, where we focus on biosignals. A biosignal refers to a measurable physiological or biological signal that conveys information about the functioning of specific physiological processes [7,33]. More specifically, our interest lies in Electrodermal Activity (EDA), which measures the skin conductance as a result of sweat gland activity [5]. Despite recent advances in the integration of biosignal data into smart environments, substantial challenges persist related to (1) the effect of wearable biosensors on smart societies at large, (2) the real-time discrepancies between subjective arousal reports and EDA signals, and (3) the extended time window of EDA signals and its consequential impact.

Existing research has shown the possibilities of continuous measuring of biosignals in the health domain [4,13,20]. However, most of this work uses biosensors in clinical or lab settings, with a few exceptions [15,36,37,41]. This contrasts the growing interest in monitoring stress and well-being in everyday life [1]. Over the last two decades, researchers have been investigating the development of unobtrusive biosensors: wearables. Wearable sensors can be used to monitor one’s physiology and behavior continuously without causing discomfort to the wearer [7]. As such, users of wearable sensors are still able to interact with their environment intuitively and naturally, while also gaining valuable insights into their health and well-being. Recent advances in wearable technology made sensors smaller, lighter, and more wearable [19,21]. Consequently, they are gaining popularity.

Incorporating wearable biosensors in things like smart homes or Intelligent Environments (IE) can allow for appropriate adaptation and early intervention, ultimately aiding in improving well-being. IE is defined as any space in our surroundings in which different aspects of the environment are controlled and adapted by intelligent agents, enhancing individuals’ experiences [3]. Here, we propose using Electrodermal Activity (EDA) signals as an enabler of early detection of long-term stress. As a biosignal, EDA allows for unbiased, real-time monitoring. Since EDA was first measured, psycho-physiological research has studied the relationship between emotional states and EDA on both a subjective (i.e. mental) and objective (i.e. physiological) level [7,34]. Together with the recent advances in wearable devices, EDA is presumed to be a measure that offers invaluable insights into the intricate interplay between the mind and body.

For more than a century, Electrodermal Activity (EDA) has been studied as a psycho-physiological measure for arousal and stress [5,33]. Consequently, there exists a comprehensive understanding of the EDA signal (for a detailed description of EDA and its features, we refer to Box 1). Empirical research has consistently shown that stress inducing situations are related to an increase in the EDA signal. To reach these conclusions, researchers have employed different methods to annotate the EDA signals. When self-reporting is utilized for annotation, it is typically conducted retrospectively. Hence, self-reported annotations and EDA signals often work on different temporal scales. Retrospective self-report relies on individuals to report their own feelings, where accurate recall difficulties is one major limitation [7,24,26]. The primary objective of this paper is to contribute to our understanding of the relationship between EDA and our subjective experience of arousal. We achieve this by evaluating EDA signals and continuous real-time self-reported arousal. Consequently, validity of EDA as a real-time stress-indicator is assessed.

The association between an increase in arousal and the corresponding elevation of EDA is often validated in short time windows, typically spanning mere seconds [39]. However, the objective of this study is to evaluate the relation between subjective arousal and EDA on larger time windows, specifically on the order of minutes. This approach holds interest as it can allow us to evaluate subjective arousal responses against EDA signals on a more detailed temporal scale while also providing valuable insights into prolonged stress exposure. Extended periods of stress can negatively influence our well-being, like experienced in bad moods [35,38]. Consequently, this study can show the potential role of EDA as a monitor of well-being.

In order to assess the viability of Electrodermal Activity (EDA) as a monitoring tool, three distinct datasets have been identified, each containing continuous EDA signals alongside continuous self-reports. When conducting an analysis that encompasses multiple datasets, some important factors need to be considered. These datasets contain recordings from diverse individuals, various contexts, and different recording devices, including both laboratory sensors and wearable technologies such as the Empatica E4. We propose a standardized methodology applicable across all datasets, aiming to aid in standardization within intelligent environments. While the inherent variability stemming from these individual differences in physiology and subjective experiences introduces complexity to the interpretation of our analyses, it also provides us the valuable opportunity to examine the robustness of our findings. This enables us to evaluate and ensure the accuracy of EDA measurements against subjective arousal.

First, we present the methodological framework in Section 2, followed by the presentation of the results in Section 3. Finally, in Section 4, we engage in a discussion of the results, their implications, and future research directions.

Fig. 1.

An idealistic electrodermal activity signal, containing three responses. Two non-specific Skin Conductance Responses (nsSCRs) and one Skin Conductance Response (SCR) as a result of the stimulus.

2.Methodology

In this study, we build upon the authors’ prior research, which involved the assessment of two Electrodermal Activity (EDA) features and self-reported arousal on a continuous scale [22]. The earlier findings were encouraging, revealing significant correlation coefficients in high-arousal scenarios. Here, we extend our study to two additional datasets to assess the generalizability of the results. Furthermore, we propose a new processing pipeline that enables us to deal with the different datasets and their different characteristics while implementing one single methodology.

2.2.Data processing

Within each dataset, two recordings are stored, self-reported arousal and an Electrodermal Activity (EDA) signal. These signals undergo discrete processing procedures, as illustrated in Fig. 2. The process ultimately leads to the computation of a correlation coefficient using both processed signals, followed by an assessment for statistical significance.

Fig. 2.

The processing pipeline for the 3 datasets. The processing of the annotation data and Electrodermal Activity (EDA) data is separated.

Table 2 provides an overview of the signal characteristics of the different datasets. It is important to note these differences because they impact signal characteristics and ultimately processing decisions. We have applied a similar method in previous work on the CASE dataset which used lab grade biosignal sensors [22]. Interestingly, the additional datasets in this work have biosignals recordings measured with the Empatica E4 wristband. These signals are inherently different because placement of the Empatica E4 EDA sensors is on the wrist, whereas the EDA sensor used for the CASE dataset was placed on the palm of the hand. It has been shown that the skin conductance on the wrist is less evident, which is clearly reflected in the range of μS in the recordings and how they differ across datasets [12,40]. This difference, together with the difference in sampling rate, is considered in the EDA processing explained in Section 2.2.2.

If we review the annotation signals in Table 2 we can observe that the recordings in the CASE and CEAP-360VR datasets are, even though distinct, rather similar. However, the K-EmoCon datasets applied a completely different method and reported continuous self-report using a method with a far lower sampling rate and resolution. These, together with the fact that recording of self-report is in retrospect, are important notes to consider when processing and interpreting the results.

Table 2

Details on the Electrodermal Activity (EDA) and arousal annotation signals and their difference across datasets

	Name	EDA recording device	Sample rate EDA (Hz)	Range EDA (μS)	Annotation recording device	Sample rate annotation (Hz)	Range annotation
[32]	CASE	Thought technology, GSR sensor	1000	0–40	Joystick	20	0.5–9.5
[42]	CEAP-360VR	Empatica E4	4	0–1.5	Joy-Con controller	10	1–9
[23]	K-EmoCon	Empatica E4	4	0–4	Likert Scale	0.2	1–5

2.2.1.Annotation processing

In the CASE and CEAP-360VR datasets, the recorded annotations are outcomes of a Cartesian-based annotation process, wherein the valence and arousal dimensions are respectively represented along the x and y axes. In the present context, only the arousal component (i.e., the y-position) is considered.

Within each dataset, distinct scales are employed to signify arousal levels. To illustrate, within the CASE dataset, the range for arousal annotation spans from 0.5 to 9.5, whereas the K-EmoCon dataset adopts a range of 1 to 5 for the same purpose, see Table 2. To normalize these ranges consistently across datasets, an adjustment is implemented, effectively rendering a neutral state of arousal ascribed to the value 0. This procedure is outlined in Algorithm 1. Subsequently, the signal is subjected to a normalization process, which corrects for interpersonal differences among participants. This entails division by the maximum arousal annotation recorded across all stimuli. This normalization approach is chosen for its ability to retain positive arousal values within the interval of [0, 1], while concurrently mapping negative arousal values onto the interval of [−1, 0] [34].

Algorithm 1:

Self-reported arousal processing

2.2.2.Electrodermal activity processing

As explained in Box 1, Electrodermal Activity (EDA) signals are typically divided into tonic activity and phasic actvity. In our study, our focus is specifically on the phasic activity, more precisely the ‘non-specific’ SCRs (nsSCRs). An nsSCR denotes any emerging SCR that lacks a direct connection to internal or external stimuli. The frequency of these nsSCRs is widely employed as a measure, particularly in contexts involving arousal, stress, and emotions [5,40]. It is even considered a more effective and valid measure compared to skin conductance level when employing both palm and wrist based EDA measures [40]. Although less frequently employed, the sum amplitude of nsSCRs also serves as a valuable measure, enabling the quantification of total ‘non-specific’ changes in EDA [5].

A challenge emerged during the analysis of EDA signals from diverse datasets. The records showed significant disparities, including variations in range, likely attributed to differences in recording devices and sensor placement. The CEAP-360VR and K-EmoCon datasets utilized the Empatica E4 wristband to measure EDA, while CASE used a sensor positioned on the index and ring finger, a common site for EDA measurement related to affective states [12]. Consequently, the EDA data within the CEAP-360VR and K-EmoCon datasets feature smaller ranges, sometimes even up to 30 times smaller compared to the CASE dataset. This discrepancy significantly impacts EDA analysis. Traditionally, a static threshold has been set for nsSCR detection, allowing for a minimum value typically between 0.01 and 0.05. However, due to the range disparities present in the raw Empatica E4 data, this conventional static threshold proved ineffective. Consequently, we propose a procedure for nsSCR detection that is independent of the recording device used.

Algorithm 2:

Electrodermal Activity (EDA) processing

Algorithm 2 provides an overview of the EDA processing. For the peak detection we made use of the neurokit2 package, which contains functions for biosignals processing. Here, we used eda_phasic() with the ‘highpass’ argument to extract the phasic activity of the signal, the highpass filter had a cutoff frequency of 0.05. The eda_phasic() function returns the phasic activity present in the signal (i.e. the skin conductance responses). To ensure that all EDA data from the different datasets can be analysed accordingly, we normalize this phasic activity. This normalized phasic signal is then passed in the eda_peaks() function which detects the nsSCRs and returns the onset, peak, and amplitude. A threshold of 0.01 μS was set for the nsSCRs amplitude, which represents the minimum EDA change to count as an elicited nsSCR [7].

Based on the location and amplitude of the nsSCRs we can create two information vectors where (1) contains a 1 where a peak is detected, and (2) contains the amplitude value at the same peak location. A moving sum filter of 20 seconds is then applied over the two information vectors, resulting in two vectors, (1) contains the nsSCRs frequency per 20 seconds and (2) contains the total nsSCRs amplitude per 20 seconds. Finally, both vectors are normalized by dividing by the maximum value recorded across all stimuli of the corresponding nsSCRs feature, see Algorithm 2.

2.2.3.Correlation and testing

Prior to performing a correlation analysis, it is essential to ensure uniform sample frequencies for both signals (i.e. the self-report signal and EDA feature signal). For all datasets, with the exception of CASE, this meant down-sampling either the self-reported arousal or the EDA signal, dependent upon which of the two has the lower sample rate. Following down-sampling, a smoothing filter is applied to both signals, employing a window size of 15 seconds.

Subsequent to these preparatory steps, the data is ready for correlation testing. The Pearson’s correlation coefficient was computed for each recording, corresponding to each participant and stimulus combination. When no nsSCRs were present in the signal, both the nsSCR frequency and amplitude signal is zero, hence no correlation value can be computed, which results in missing values. We hypothesise a positive Pearson’s correlation coefficient between the continuously self-reported arousal and nsSCRs’ frequency and amplitude. In other words, we expect an real-time increase in nsSCRs’ frequency and amplitude when self-reported arousal is increased.

We then performed statistical tests to assess the statistical significance of the correlation values. For our CASE and CEAP-360VR datasets we encountered 8 distinct conditions, corresponding to 8 distinct videos. Consequently, we obtained on average 8 observations per participant per nsSCR feature, although it is noteworthy that some participants had missing correlation measures. Hence, for the CASE and CEAP-360VR datasets we conducted a multilevel (ML) analysis, also known as a mixed-effects or mixed linear model [16]. ML analysis is capable of handling such missing data. More details on ML analysis is given in the following section. Note that correlation results for K-EmoCon do not contain repeated measures and thus no ML modeling is applied, we have just one data point (i.e. correlation value) per participant. Hence, we test the correlations on their statistical significance against the null hypothesis by employing an one-sampled t-test.

2.2.4.Multilevel analysis

We employ multilevel (ML) analysis because of the nested structure of our observations within participants. This choice becomes particularly relevant when considering the potential influence of individual differences in biosignals data. Individual participants inherently possess unique physiological and psychological traits, contributing to variations in their biosignals, including Electrodermal Activity (EDA). For instance, the number of sweat glances on the skin differs from one individual to another [7]. These differences are expected to have implications for the monitoring and analysis of affective states through biosignals. In addition to these physiological variations, we anticipate that individuals possess differing abilities to interpret their psychological states and, consequently, may annotate their experiences differently. For instance, people can differ in their tendency to adhere to socially desirable responding [24]. More generally, the interpretation of arousal (and valence) differs across persons and circumstances [17]. Hence, we present our multilevel model, defined as

(1)OutcomeVariableij=B0j+B1j·Arousal1ij+eij,

where OutcomeVariableij is a representative for our two outcome variables, which represent the correlation values of amplitude and frequency of nsSCRs against self-report. Arousal1ij reflects the level of arousal depicted in the stimuli which is either low (0) or high (1). The subscript i refers to the lower level 1 (i.e. repeated measure of video conditions) and the subscript j to level 2 (i.e. the individual participants). Multilevel analysis allows to compare the individual participants since each participant will have its own slope B1j and intercept B0j. It shows the variance across participants by looking at individual slopes and intercepts. It also allows to evaluate the overall or average effects by looking at the fixed effect. To provide more detail, we can rewrite the coefficients of the model as

(2)B0j=γ00+U0j andB1j=γ10+U1j,

where γ00 is the overall mean intercept and U0j j’s deviation from the mean intercept, and γ10 the overall mean slope and U1j j’s deviation from the mean slope. To be able to elaborate on the significance of the model, we first build a baseline model, also called a random effect model, defined as

(3)OutcomeVariableij=B0j+eij,

here we only set the cluster variable (i.e. the participants) and calculate the mean intercepts, both within and between participants. We then build on top of this baseline model and test the effect of arousal level and whether the fit of the model improves. Multilevel analysis was performed in R version 4.3.1, using the lme4, lmerTest, and effects packages.

3.Results

Illustrated in Fig. 3 are three representative examples showcasing the relation between self-reported annotations and Electrodermal Activity (EDA) features within the context of a high arousal video. These instances are drawn from discrete datasets, and notably, the examples 3a and 3b affirm the hypothesis. Nonetheless, it is essential to underscore that these examples have a limited scope as they show participants which were exposed to stimuli eliciting heightened arousal, except for Fig. 3c.

Fig. 3.

Example of the processed self-reported arousal against the processed ‘non-specific’ Skin Conductance Response’s (nsSCR) frequency and amplitude, given over the duration of one video with a high arousal context, extracted from the (A) CASE, (B) CEAP-360VR, and (C) K-EmoCon dataset.

First we calculated the Pearson correlation coefficients, which in rare cases resulted in a missing value. This is due to the absence of a nsSCR, hence one of the vectors has a constant value of zero. With the CASE dataset 5% of correlation coefficients was missing. The CEAP-360VR correlation coefficients contained 1.5% missing values. Finally, the correlation coefficient calculated from the K-EmoCon dataset contained no missing values.

Figure 4 presents the calculated correlation coefficients for different arousal contexts, across datasets. Notably is the congruence between the findings from the CASE dataset in previously established research [22], while here we used a different nsSCRS detection method. Furthermore, Fig. 4 showcases similar results for the CASE and CEAP-360VR datasets. Despite the distinct differences between the datasets, encompassing different participants, contextual variables, and instrumentation, it is noteworthy that for both the correlation coefficients show elevations in instances of heightened arousal. Table 3 provides an overview of the exact median values across the datasets and arousal conditions.

Fig. 4.

Distribution of correlation coefficients for frequency and amplitude of ‘non-specific’ Skin Conductance Responses (nsSCRs) with self-reported arousal. It shows the results of all three datasets and accordingly on the different arousal conditions present in the datasets.

4.Discussion

To create a truly smart society, we believe that next to efficiency and effectiveness, there is great value in being able to ‘sense’ an individual’s health and well-being [18]. Having a human understanding allows for appropriate adaptation of the environment, creating a loop of information gathering and action, adding to a smart society. An example of this is the ability to identify affective states of hospitalized children, providing appropriate adjusted surroundings such as offering soothing stimuli to relieve the child’s anxiety. Or, adaptive lighting of the room based on affective state of the inhabitant. This personalized and responsive approach to environments showcases the potential of integrating sensory technologies into our society.

Pro-longed experience of stress is reflected in bad moods and a decrease in performance and well-being [14,35]. Being able to detect these stress levels can allow for appropriate intervening through smart environments and support individuals, functioning as a first step in supporting the complex construct well-being [9]. Recent advances in wearable biosensors introduce new directions to monitor these states. One particular biosignal that can be measured trough wearable devices is Electrodermal Activity (EDA). We assessed EDA’s potential as a real-time indicator of stress levels. We did so by conducting an analysis on three datasets (i.e. CASE [32], CEAP-360VR [42], and K-EmoCon [23]) containing both continuous EDA signals and continuous self-reported arousal. We calculated the correlation between EDA features, namely non-specific Skin Conductance Response (nsSCR) amplitude and frequency, and self-reported arousal. In the CASE and CEAP-360VR datasets, our analysis revealed significant positive correlations, where in the K-EmoCon dataset these were absent. Next, we will consider these results along the following discussion points: (1) EDA as a real-time indicator, (2) extended time windows, and (3) the ecological validity. Subsequently, we present the limitations of the study and end with a conclusion.

The results showcase that EDA is capable of functioning as a real-time measure for arousal. The positive correlations in high arousal situations, obtained from CASE and CEAP-360VR, underscore that as subjective arousal increases, EDA features also exhibit increase across a continuous temporal scale. For over a century different theories have been developed in relation to the subjective experience and physiological states. One of these theories is the James–Lange theory which states that our affective experiences, such as emotions, are a result of perceiving to physiological changes in our body, for example we feel fear or joy after we notice our heart rate increases or our palms sweat [11]. An alternative theory is the Schachter and Singer’s two-factor theory of emotion, which states that affective states are determined by a combination of both physiological responses and the cognitive appraisal of the situation. This means that the same physiological state can lead to different experiences, depending on the cognitive interpretation [30]. Although our research is not able to validate either of such theories, which is very complex as pointed out by Cacioppo and Tassinary [6], it is interesting that the results of two datasets suggest the mapping between physiological changes and subjective experience is there, on a continuous scale, indicating the two are indeed interconnected. Note that this mapping is complex and not universal (i.e. the same across individuals) nor constant (i.e., the same at different instances for a given person) [14]. It is regulated by our cognitive processes of evaluation, which may be one the reason our third dataset, K-EmoCon, does not show this mapping.

In contrast to traditional practices in EDA validation studies, which often focus on time windows no longer than a few seconds [39], our study employed extended time windows (i.e., minutes). This approach enables us to look at the EDA signal differently and extract different features. Instead of looking at acute responses, we are interested in prolonged diffused responses. As a results we look at different features of the signal. Our decision to look at the nsSCR as our measure is one of them, which has been validated in other work [40]. The extended time windows and different features also means we can examine different psychological constructs, including moods, prolonged stress, and, ultimately, well-being. Summarizing, we are not using EDA traditionally as for acute stress responses, but rather as underlying continuous stress indicator. The presence of positive and significant correlation coefficients in high arousal scenarios underscores the potential for gaining insight into stress dynamics and implementing timely interventions, particularly when high arousal persists over extended duration.

The CEAP-360VR and the K-EmoCon datasets featured EDA measurements obtained via a wearable device, specifically the Empatica E4. In contrast, EDA measures for the CASE dataset were recorded in controlled lab environments using laboratory sensors. This means that EDA recordings in the CASE dataset are significantly different from the CEAP-360VR and K-EmoCon dataset. For one, wearable devices measure EDA on the wrist, whereas laboratory sensors measure on the palm of the hand, which is recommended [12]. It is knows that the phasic signals originating from the wrist may be less pronounced compared to those originating from the fingers or the palmar surface [40]. Overall, the signal quality of wearable devices tends to be lower compared to lab-grade devices, partly due to the use of dry sensors [2]. Despite these important differences, it is noteworthy that both the CASE and CEAP-360VR datasets yielded similar outcomes in cases of high arousal, suggesting applicability of Empatica E4 in this context. This finding bears particular significance concerning a smart society, which cannot rely on laboratory sensors but must incorporate wearable devices. Consequently, the results from our study advocate for the integration of wearable devices into our living environments to provide individual support. This integration can assist in timely interventions when stress levels escalate, ultimately contributing to improved well-being. Note that K-EmoCon dataset does not show the same results while employing an Empatica E4, and thus not support the argument given. We believe there are several factors that can explain this deviance, provided in the following paragraph which present the first of several limitations that warrant consideration: (1) the challenge of generalization, (2) constraints related to low arousal states, and (3) individual variations in self-report.

Firstly, while the different datasets facilitated an extensive examination of arousal within varied contexts, both in controlled laboratory settings and more realistic scenarios, it is essential to acknowledge that they may not fully capture the intricacies of real-world situations. This complexity poses challenges in translating our findings into practical real-world applications and generalizing them effectively. One step towards real-world applications is the use of wearable devices to measure Electrodermal Activity (EDA). Two datasets, namely CEAP-360VR and K-EmoCon, did employ such wearable devices. Consequently, the outcomes derived from these datasets hold significant relevance concerning their potential applicability to real-world monitoring scenarios. However, both datasets yielded diverse results, making it hard to draw definitive conclusions from these findings. The results from K-EmoCon did not support our hypothesis. It is noteworthy that this dataset is designed to minimize intense emotional responses. Additionally, self-report data in this dataset suffered from retrospective reporting and poor resolution. Prior research has highlighted the limitations associated with such retrospective annotations and the time gap it introduces between the actual affective state and its annotation [24]. This may explain the deviance with our hypothesis in the K-EmoCon dataset. Overall, to support our findings, future research should be focused on expanding the variety of datasets and contexts in which data is gathered. Furthermore, long-term studies would be beneficial for the assessment of EDA in real-world application for smart societies.

Secondly, our study consistently identified significant effects between EDA features and self-reported arousal in instances of heightened arousal. Notably, as arousal levels diminished, correlation coefficients exhibited a corresponding decrease. Consequently, our findings do not offer comprehensive insights into states of low arousal, which is also limiting their direct applicability in real-world contexts. It suggests the existence of a threshold, beyond which physiological indicators directly reflect arousal levels. As such, accounting for the anticipated level of arousal is crucial when utilizing nsSCR frequency and amplitude as indicators. Future research should aim to precisely determine this threshold.

Finally, an examination of the spread of correlation coefficients, even in high arousal scenarios, reveals notable variance. This variance may be attributed to individual differences in reflective abilities, such as an individuals’ capability for annotating and articulating their physiological states. It is possible that not all individuals possess the same degree of competence. There are issues to consider in self-report such as socially desirable responding (SDR), acquiescent responding (AR), and extreme responding (ER) [24]. People can differ in their tendency to adhere to SDR and present self-report that reflects what they think the researches expect from the stimuli, rather than reflect what they are feeling. This tendency can be linked to personality, but since we do not have personality information on the participants in the datasets, we were not able to test for this. AR, which is a term used for people who tend to agree, can be viewed as an individual difference linked to conformity and impulsiveness [24]. Again, we lack the information on personality to check for this effect. ER, which has been validated as a consistent individual difference, where ambiguity, emotional arousal, and rapid responding induce temporary increases in extreme responding [24]. This makes ER also a likely factor to consider. With normalization we try to control for this factors, but in cases where someone is experiencing little arousal yet annotates extremely over all stimuli, the normalization does not suffice. Finally, people simply do not have an infinite ability to recall and reflect all relevant information [24]. They can be overwhelmed with information, in this case rapid changes in stimuli. This can make the task of annotation difficult, for some more than for others.

Collecting physiological data and using it in a smart society brings a responsibility regarding safety and privacy of this sensitive information. For one, users of wearable devices should have ownership of their data. Additional concerns are re-identification and linkage risks resulting from routine activities and regular movements recorded through the device. A potential solution involves implementing anonymous user IDs to communicate across devices. Furthermore, incorporating local real-time processing on the device (i.e. federated learning) can address privacy and safety concerns. Users should be notified of any policy changes to ensure transparency and user awareness [8]. Finally, it’s also important to acknowledge that self-tracking, while beneficial, can escalate into an obsession, potentially causing more harm than good. Further research is needed to determine how to maintain a balanced approach to wearable devices integrated in our society [8].

In conclusion, our findings reveal a significant correlation between EDA features and self-reported arousal within contexts characterized by high arousal levels. This consistency in results holds across two distinct datasets (i.e., CASE and CEAP-360VR), with the third dataset (i.e., K-EmoCon) deviating. This divergence can arguably be attributed to inherent differences from the other datasets, such as the retrospective annotation as well as the frequency and resolution of the self-report. Our research underscores the possibilities of continuous monitoring of EDA signals, particularly in real-time reflection of subjective arousal under high arousal conditions. The results highlight the potential of EDA monitoring through wearable devices as a promising avenue for real-world stress and arousal detection. Overall, this study contributes to our understanding and monitoring of physiological responses, adding to the larger vision of creating a smart society that can adapt to individual needs. As such, smart societies may one day be able to be truly smart and support individuals in their pursuit of well-being.