Challenges in Using Skin Conductance Responses for Assessments of Information Worker Productivity

Anam Ahmad, Georgia Institute of Technology, United States, anam@gatech.edu
Thomas Ploetz, Georgia Institute of Technology, United States, thomas.ploetz@gatech.edu

DOI: https://doi.org/10.1145/3594738.3611360
ISWC '23: Proceedings of the 2023 International Symposium on Wearable Computers, Cancun, Quintana Roo, Mexico, October 2023

Breaks as discretionary self-interruptions can have beneficial effects on information worker productivity and well-being. This has design implications for potential productivity tools that can assess opportune moments to suggest these breaks. Electrodermal Activity (EDA) is a good psychophysiological metric to capture changes in autonomic activity resulting from affective states that necessitate breaks. Wrist-worn sensing platforms have been heralded as effective means for EDA-based affective state assessments in real-life scenarios. However, our study finds no correlation even in a controlled setting with a constrained operational definition of productivity and well-researched EDA measurement and processing techniques. We reflect on our rationale against prior success reported in laboratory and ambulatory assessments of EDA.

Keywords: information work; productivity; cognitive load; interruptions; electrodermal activity; wearable EDA sensors

ACM Reference Format:
Anam Ahmad and Thomas Ploetz. 2023. Challenges in Using Skin Conductance Responses for Assessments of Information Worker Productivity. In Proceedings of the 2023 International Symposium on Wearable Computers (ISWC '23), October 08--12, 2023, Cancun, Quintana Roo, Mexico. ACM, New York, NY, USA 6 Pages. https://doi.org/10.1145/3594738.3611360

1 INTRODUCTION

Information workers typically multitask due to frequent external and self-interruptions [27]. Several studies have concluded that external interruptions can be detrimental to the quality of work and personal well-being [3, 5]. However, discretionary self-interruptions in the form of short breaks can be beneficial for performance, specifically in finding new solutions and reducing impasses in creative problem-solving tasks [6]. They can also alleviate stress and have positive consequences on overall well-being [27]. These beneficial short breaks when characterized as self-interruptions, are often ascribed to fatigue, frustration, or boredom [9, 26] that can be measured as affective states [35].

Several systems have been developed with the aim to predict opportune moments for breaks which can have a restorative effect on cognitive resources. To that end, wearable sensors are widely used to measure cognitive load through physiological signal monitoring. Common signals are Heart Rate Variability (HRV), Electrodermal Activity (EDA), Skin Temperature (ST), and Accelerometry [1, 8, 15, 17, 28] in addition to more intrusive modalities such as Electroencephalography (EEG) [46] and Pupillometry [25, 28]. Recording this data beyond laboratory settings has been facilitated by growing popularity of commercially available wearables like the Empatica E4 and Embrace wristbands (Empatica Srl, Milan, Italy). EDA is considered a good biomarker for psychophysiological arousal that underlies affective states [11, 17], hence featuring in these wearable sensors. HRV is also a popular modality, but it is affected by both branches of the autonomic nervous system (sympathetic and parasympathetic) compared to EDA which is purely sympathetic [11].

Studies that assess continuous monitoring of affective states under cognitive load often administer computer-based tasks like the Stroop test [1, 8, 12, 36, 43], observe participants in learning environments (i.e., a classroom) [15, 17] or in interactive scenarios [12, 37, 38]. As a step towards investigating this in information work, this paper explores if EDA would be feasible for laboratory (and eventually ambulatory) scenarios for detecting loss of productivity, thus characterizing opportune moments for suggesting breaks.

We conducted an experiment where participants attempt a word-generation task with tight constraints, and controlled for the opportunity to self-interrupt and switch to another such task, making the break "short". We defined a typology for productive and non-productive condition classes for the scope of the experiment, with the break-opportune condition being the extreme non-productive class. For the first iteration of the study, we collected EDA using wrist-worn sensors but transitioned to toe measurements to improve signal quality. We verified that the EDA features extracted for observation windows were significantly different for productive vs non-productive conditions, but found poor classification accuracy for our trained model beyond a simple binary assessment, rendering wearable EDA-based productivity assessments challenging if not impossible. We share our rationale and reflections for extending automated affect detection for information worker well-being.

2 BACKGROUND AND RELATED WORK

Automated Affect Assessment: Information work often consists of goal-directed tasks. A low-cost option to deliver interruptions is at moments of lower mental workload that occur between subtask boundaries [4], but this necessitates prior knowledge of the task structure. Changes in mental workload correspond to an imbalance between challenge and skill [13] that manifests as changes in physiology [32]. Thus, physiological monitoring can be used to automatically (and potentially in real time) reason about when to suggest breaks [28]. Similar work has been done for automated affect assessment of students’ engagement in classroom settings [14, 15, 17], although "engagement" is explored as a multidimensional construct comprising of emotional, behavioural and cognitive engagement [17], the latter being most related to our work. Another pertinent setting wherein cognitive load and engagement is measured is social interaction in parent-child dyadic groups [12, 21, 37]. Wrist-worn EDA sensing is common in all these studies due to its popularity as an unobtrusive modality that can pick up changes in humans’ affective states. Boredom and frustration are the states relevant to our experiment.

Electrodermal Activity (EDA). EDA assessment involves measuring skin conductance via two electrodes in units of micro-Siemens (μS). This peripheral skin property changes with the innervation of eccrine sweat glands [10] due to sympathetic (flight or fight) autonomic nervous responses. EDA can be decomposed into two components: a slow-moving tonic signal called Skin Conductance Level (SCL), and a fast-changing phasic response called Skin Conductance Response (SCR) that captures the morphology of peaks. SCRs can further be categorized into two types: event-related SCRs (ER.SCR) that occur in response to external stimuli, and non-specific SCRs (NS.SCR) which occur in the absence of such apparent external stimuli. There are various algorithms that can be used to decompose EDA into its components [7, 19, 22] and to count the number of SCR peaks [30, 39]. The experiments in this study were conducted in a controlled environment with no apparent external stimuli, making the observed SCR peaks NS.SCRs.

Critique of EDA Assessment. Traditional electrode configurations of EDA measure from the fingertips, which is also the case for gold-standard measurement devices like Flexcomp Infiniti [47], MindWare EDA [33], Biopac MP36R [11, 24]). However, this configuration impedes activity conditions in computer-based tasks popular in studies for cognitive load assessment. These devices are also not portable and are quite expensive. Thus, various studies leverage EDA collected from cheaper, commercially available wrist-worn sensors like Affectiva Q, iCalm, Empatica E4 due to ease of use in ambulatory settings and the ubiquitous form factor of a watch. This does not come without its challenges. Despite several methods explored for minimizing artifacts in ambulatory EDA measurements [12, 18, 31], there exists conflicting evidence for EDA's feasibility for affect detection for scenarios described earlier. Hernandez et al. [21] found that wrist-worn EDA accurately captured engagement in social interactions but discarded 31% of data due to artifacts. Another study [37] discarded 73% of data, concluding the wrist location as unsuitable. In the classroom setting, Gao et al. [17] reported weaker EDA correlations for cognitive engagement, while DiSalvo et al. [15] found no correlations with EDA features. Many studies use 0.05μS as the lower bound for valid EDA [12, 18, 31] based on the seminal publication guidelines for EDA measurements [16] but this value is a threshold for SCR peak amplitudes, not raw EDA. Milstein et al consider raw EDA below 0.5μS as noise [37] as per the original E4 documentation and the gold-standard EDA device manufacturer Biopac outlines 1-20μS as the reference range [24]. We thorougly explored this discrepancy in our experiments.

3 CASE STUDY: EDA FOR PRODUCTIVITY

We sought to examine EDA's effectiveness as a modality for identifying lack of productivity, which could characterize the right time to suggest a break. We designed a computer-administered experiment with word-generation tasks, controlling for opportunities where participants could self-interrupt out of unproductive states. We chose word-based tasks inspired by Stroop tests from the literature [1, 8, 12, 36, 43], but made it loosely goal-directed to artificially represent daily information work. EDA from the distal forearm (wrist) and toe-tips was collected in two iterations of the experiment.

Figure 1
Figure 1: Bitalino EDA sensor w/ toe electrode configuration

3.1 Experimental Setup

Our participant pool comprised of 35 (20 for first iteration, 15 for the second) university students majoring in computer science and 2 software engineers 18-30 years of age, fluent in English. The experiment was conducted as a voluntary study after obtaining informed consent via approved IRB forms. The experiment consisted of a single session of 45-60 minutes in two phases, where participants were given a 9-letter scrambled anagram in a Google doc and asked to list as many valid English 4-letter words from it as possible without repeating any letters. In the first phase they were given only one such anagram "PWIOREKAY", following which they were given two anagrams "BFIOJERAC", "PVIOXELAM" in separate Google docs. All 3 anagrams are of similar difficulty and adopted from [48]. For both phases, they were instructed to inform the experimenter when they could not think of any more words, bringing the phase to an end. Participants could choose to switch back and forth between the two anagrams in the second phase, but they were not informed of the same to observe if they made this choice autonomously. At the end, participants completed a post-activity metacognitive questionnaire reflecting on their choices.

3.2 Data Collection

3.2.1 EDA Data. In the first iteration of the study, we used the Empatica E4 to collect EDA data. It measures skin conductance across two points on the wrist and has the familiar form factor of a wrist-watch, making it a convenient wearable sensor. However, we noticed that the recorded measurements were much below the reference range for human skin conductance, with flat readings below 0.5μS for 17 out of 20 participants. To verify whether this was due to sensor wear/tear or defect, we measured data from the fingertips on the same hand and observed values consistently within the reference range. We concluded that–contrary to popular belief–the wrist was actually not the best measurement location for the study, but measuring from the fingertips would impede the activity and would also be difficult to eventually deploy.

Based on experiments correlating different on-body locations for good EDA measurement [40], we amended the study to incorporate the Bitalino EDA sensor with foot-based electrodes (Figure 1). Participants cleaned two toes of the left foot with a wipe then secured the electrodes with silicone protectors. This required additional signal processing since this was a rudimentary sensor susceptible to electrical noise, but the reference range in non-activity conditions was correct (>1 μS). The setup was not disruptive to the activity, albeit not as convenient to "wear" as the E4.

3.2.2 Activity Annotations. Objectively defining productivity for information work is challenging. However, our experiment involved a tightly constrained stimulus-limited task adopted from studies about creative problem-solving [48]. The participants produced words in Google docs that offer version control at roughly a 60-second granularity, hence our observation windows were 60 seconds long. This is enough to capture SCR peaks in their entirety from rise to recovery [18]. Since we were interested in characterizing periods opportune for suggesting breaks, our annotation procedure was based on identifying unproductive conditions manifesting as sustained inactivity. Therefore, we defined productivity as non-zero word count, with the goal to identify unproductive periods. Crucially, inactive engagement was not considered unproductive wherein participants may have a zero word count but are actively engaging in thinking about possible words. We called these "reading the question" conditions (1 and 5) where participants could either gain ideas to resume the task or get stuck thinking for too long. We argue that the latter is an opportune moment to recommend a break. Based on our task complexity and observing participants, 2 or more consecutive minutes was "too long" for our study. The typology of all conditions is described in Table 1.

Table 1: Activity categories: Conditions are described using tasks i, j and word count w in 60s consecutive observation windows t, t + 1. Condition 4 is most opportune for a break.
Condition: Description Remarks
1: Ø → it + 1 Starting the task
2: it → it + 1, wt = 0, wt + 1 ≠ 0 Got new ideas, regained focus
3: it → it + 1, wt ≠ 0, wt + 1 ≠ 0 Productive, continued on task
4: it → it + 1, wt = 0, wt + 1 = 0 Stuck/unproductive
5: it → it + 1, wt ≠ 0, wt + 1 = 0 May become unproductive soon
6: it → jt + 1, wt = 0, wt + 1 ≠ 0 Switch productive
7: it → jt + 1, wt = 0, wt + 1 = 0 Switch unproductive

3.3 Data Analysis

We analysed the collected EDA data and the activity annotations to find correlations – effectively replicating protocols from previous work. The methods used are detailed in this section.

3.3.1 EDA Preprocessing. We first trimmed the raw EDA based on the synchronization markers for Phase 1 and Phase 2. To remove high-frequency noise, we used a 3rd order low-pass Butterworth filter with cutoff frequency of 0.35Hz [29]. We then decomposed the raw EDA signal into its tonic and phasic components using the cvxEDA method [19] provided by the NeuroKit 2 physiological analysis toolkit [34]. The phasic component comprises of the SCR peaks that we subsequently leveraged for knowledge-based features. We used the peak counting method outlined by Kim et al [30] since it is robust to artifacts.

3.3.2 Feature Extraction and Classification. We divided the preprocessed EDA data into 60-second segments with a corresponding "condition" label (1-7) from the activity annotation and used a 60-second sliding window with no overlap to extract features for a classification back-end. The following knowledge-based features were extracted for each window similar to previous work [15, 23, 49]: i) mean and standard deviation of the raw EDA values, ii) mean and standard deviation of the phasic signal, iii) mean and standard deviation of the tonic signal, iv) number of scr peaks, v) average scr peak amplitude, and vi) average scr rise time. These features have been validated as signs of general arousal [14, 15]. The most related previous work found that even their best classification model (a Random Forest classifier) performed poorly [15]. Other models they tried were Naive Bayes, K-NN and Support Vector Machines. Also, since most of the chosen features are partially derived from other included features, we only used the Random Forest classifier for this study.

3.4 Evaluation

3.4.1 Statistical Tests. To assess the quality and explainability of our features, we tested whether i) EDA is significantly different in productive vs. unproductive windows. Since the knowledge-based features are signs of general arousal, we expected value differences to emerge in productive vs non-productive windows. To assess whether self-interruptions affected EDA, we tested if ii) EDA is significantly different in productive windows in phase 1 vs phase 2. We employed Wilcoxon rank-sum tests for each feature for the most extreme conditions of each type (productive: condition 3, unproductive: condition 4). We performed these tests before feeding the features to our classifier to confirm if they were characterizing the differences in productive vs unproductive conditions to a significant extent.

3.4.2 Leave One Subject Out (LOSO). Time-domain features of EDA (tonic / phasic) have high inter-subject variability [44], hence the signals are z-score standardized before decomposition [34]. This reduced variability allowed us to conduct a LOSO evaluation, a step towards training a large-scale generalizable model. Phase 2 data was used to incorporate instances from all possible ground truth conditions. F1 (macro) and accuracy scores were calculated for each held out participant, and averaged across all participants. 7-class classification is a challenging setting with limited data, so we also performed binary classification by collapsing conditions 1, 2, 3 and 6 into a "productive" class and 4, 5 and 7 into "unproductive". This deviates from practical utility and does not allow for nuance in the definition of productivity, but would offer feasibility insights.

4 RESULTS

As mentioned in Sec. 3.2.1, 17/20 participants exhibited a flat EDA signal with values consistently below 0.5μS from the wrist, which is below the reference range. This added evidence to the critique from literature that wrist EDA recordings are unreliable [2]. Moreover, classifier performance only reached an average F1 score of 0.14 across the 7 classes and 0.42 for the binary classification. We thus focus on detailing results from the foot EDA experiments. The final dataset consisted of data from 9/15 participants. We had to discard 30% of our data because of artifacts like loss of contact and anomalous peaks resulting from touching a conductive surface. Some of these limitations in our rudimentary sensor setup can be overcome with a wireless data-streaming module.

4.1 Statistical Correlation Tests

For the first test (Table 2), 4 knowledge-based features supported that EDA was different in productive compared to unproductive windows. Surprisingly, the number of peaks, which is considered a direct measure of arousal, was not found to be a discriminating feature. For the second test, 6 knowledge-based features supported that EDA was different in phase 1 compared to phase 2, wherein participants had the opportunity to self-interrupt.

Table 2: Test results for knowledge-based features, *p < 0.01
Test Features
i) eda_mean*, eda_std, eda_auc*, scr_count, scl_mean*, scl_std, scl_auc, scr_avg_peak_amplitude, scr_avg_rise_time, phasic_mean, phasic_std, phasic_auc, scl_mean, scl_std, scl_auc*
ii) eda_mean, eda_std, eda_auc*, scr_count, scl_mean, scl_std, scl_auc*, scr_avg_peak_amplitude*, scr_avg_rise_time, phasic_mean*, phasic_std*, phasic_auc*, scl_mean, scl_std, scl_auc
Figure 2
Figure 2: Class-relative performance for LOSO evaluation.

4.2 Leave One Subject Out (LOSO)

The f1 score averaged across LOSO run of each participant was 0.184 (SD=0.13). Figure 2 summarizes the class-relative performance breakdown consolidated across these runs, wherein the test samples are aggregated across all participants. Most windows were misclassified as condition 3 and a fraction of conditions 5 and 6 were classified correctly. We also experimented with a latent representation using ECDF features [20, 41] but found no improvement. Our data was imbalanced (1: 6%, 2: 8%, 3: 42%, 4: 5%, 5: 15%, 6: 21%, 7: 3%) but even the most frequently occurring condition 3 had an average f1 score of 0.43 (SD=0.15) across all participants. None of the extreme unproductive cases (condition 4) were classified correctly.

4.3 Discussion

Our feature comparison tests revealed significant differences between the extreme conditions (3 and 4), but the model could not accurately classify all 7 conditions. A potential intervention system for predicting opportune moments for breaks must reason at a granularity finer than just a binary one. For instance, the system should be able to distinguish between inactive engagement and sustained inactivity to make judgements accordingly. However, to find evidence for feasibility, we also performed binary classification as described in Sec. 3.4.2. The overall average f1 score was 0.46 (0.85 for "productive", 0.13 for "unproductive) – which is not promising. While this imbalance can be improved for training, opportune moments for breaks are likely to be a fraction of the total work day anyway. To assess the affect impact, we analysed a post-study meta-cognitive questionnaire completed by the participants. 67.5% took the opportunity to switch between anagrams, primarily when they could not think of words but also when they wanted to try something new. On average, they exhibited a better word production rate (1.91 wpm) than the ones who did not switch (1.54 wpm). But we suspect that this mixed intent blurred the difference between boredom and engagement, blunting the discriminative power of the peak count (scr_count) feature. Additionally, the tasks were not challenging enough to elicit frustration in every participant.

This study aligns with nearly all best practices suggested for EDA research in the HCI community [2], except we did not control for caffeine intake and medication. Also, our choice of using z-scores to tackle inter-subject variability differed from approaches that found more success using min-max range corrections [15, 21, 43]. The rationale behind not using min-max was that we should not just use smallest and largest values from the experiment, but the minimum and maximum values that participant can exhibit – warranting separate measurements. Literature cautions about the reliability of such range corrections [10] and deems z-scores more reliable. Making recognition models generalize to unseen participants is challenging [42], but more meaningful for deployment [15].

5 CONCLUSION

We aimed to assess whether a person's EDA could characterize the affective states underlying loss of productivity in information work. We sought evidence for any correlations in laboratory settings, hence we chose a tightly constrained stimulus-limited computer-based task for our experiments. We first tried measuring EDA from the wrist, and found evidence that toes would be a better location for collecting data. After comparison tests establishing significant differences in EDA between the most typical productive and unproductive conditions, we trained a classifier, which was not able to accurately discriminate among the full array of conditions.

Without consensus on the measurement location, processing techniques and validated ranges, ambulatory measurements present bigger challenges. Some studies use frequency-domain features from finger-based EDA [44, 45], but these need to be explored further with data from more convenient measurement sites.

ACKNOWLEDGMENTS

We would like to thank Sashank Varma for guiding us with the experiment design, Alexander T Adams for helping us with the foot-based EDA sensor setup, the participants who volunteered for our study, and the anonymous reviewers for their valuable constructive feedback.

REFERENCES

  • Ane Alberdi, Asier Aztiria, and Adrian Basarab. 2016. Towards an automatic early stress recognition system for office environments based on multimodal measurements: A review. Journal of biomedical informatics 59 (2016), 49–75.
  • Ebrahim Babaei, Benjamin Tag, Tilman Dingler, and Eduardo Velloso. 2021. A Critique of Electrodermal Activity Practices at CHI. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Yokohama, Japan) (CHI ’21). Association for Computing Machinery, New York, NY, USA, Article 177, 14 pages. https://doi.org/10.1145/3411764.3445370
  • Anja Baethge and Thomas Rigotti. 2013. Interruptions to workflow: Their relationship with irritation and satisfaction with performance, and the mediating roles of time pressure and mental demands. Work & Stress 27, 1 (2013), 43–63.
  • Brian P Bailey and Shamsi T Iqbal. 2008. Understanding changes in mental workload during execution of goal-directed tasks and its application for interruption management. ACM Transactions on Computer-Human Interaction (TOCHI) 14, 4 (2008), 1–28.
  • Brian P Bailey and Joseph A Konstan. 2006. On the need for attention-aware systems: Measuring effects of interruption on task performance, error rate, and affective state. Computers in human behavior 22, 4 (2006), 685–708.
  • Flora Beeftink, Wendelien Van Eerde, and Christel G Rutte. 2008. The effect of interruptions and breaks on insight and impasses: Do you need a break right now?Creativity Research Journal 20, 4 (2008), 358–364.
  • Mathias Benedek and Christian Kaernbach. 2010. Decomposition of skin conductance data by means of nonnegative deconvolution. psychophysiology 47, 4 (2010), 647–658.
  • Javad Birjandtalab, Diana Cogan, Maziyar Baran Pouyan, and Mehrdad Nourani. 2016. A Non-EEG Biosignals Dataset for Assessment and Visualization of Neurological Status. In 2016 IEEE International Workshop on Signal Processing Systems (SiPS). 110–114. https://doi.org/10.1109/SiPS.2016.27
  • Christine Bosch and Sabine Sonnentag. 2019. Should I take a break? A daily reconstruction study on predicting micro-breaks at work.International Journal of Stress Management 26, 4 (2019), 378.
  • Wolfram Boucsein. 2012. Electrodermal activity. Springer Science & Business Media.
  • Jason J Braithwaite, Derrick G Watson, Robert Jones, and Mickey Rowe. 2013. A guide for analysing electrodermal activity (EDA) & skin conductance responses (SCRs) for psychological experiments. Psychophysiology 49, 1 (2013), 1017–1034.
  • Donna L Coffman, Xizhen Cai, Runze Li, and Noelle R Leonard. 2020. Challenges and opportunities in collecting and modeling ambulatory electrodermal activity data. JMIR biomedical engineering 5, 1 (2020), e17106.
  • Mihaly Csikszentmihalyi, Mihaly Csikszentmihalyi, Sami Abuhamdeh, and Jeanne Nakamura. 2014. Flow. Flow and the foundations of positive psychology: The collected works of Mihaly Csikszentmihalyi (2014), 227–238.
  • Elena Di Lascio, Shkurta Gashi, and Silvia Santini. 2018. Unobtrusive assessment of students’ emotional engagement during lectures using electrodermal activity sensors. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 2, 3 (2018), 1–21.
  • Betsy DiSalvo, Dheeraj Bandaru, Qiaosi Wang, Hong Li, and Thomas Plötz. 2022. Reading the Room: Automated, Momentary Assessment of Student Engagement in the Classroom: Are We There Yet?Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 6, 3 (2022), 1–26.
  • Society for Psychophysiological Research Ad Hoc Committee on Electrodermal Measures, Wolfram Boucsein, Don C Fowles, Sverre Grimnes, Gershon Ben-Shakhar, Walton T Roth, Michael E Dawson, and Diane L Filion. 2012. Publication recommendations for electrodermal measurements. Psychophysiology 49, 8 (2012), 1017–1034.
  • Nan Gao, Wei Shao, Mohammad Saiedur Rahaman, and Flora D Salim. 2020. n-gage: Predicting in-class emotional, behavioural and cognitive engagement in the wild. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 4, 3 (2020), 1–26.
  • Shkurta Gashi, Elena Di Lascio, Bianca Stancu, Vedant Das Swain, Varun Mishra, Martin Gjoreski, and Silvia Santini. 2020. Detection of artifacts in ambulatory electrodermal activity data. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 4, 2 (2020), 1–31.
  • Alberto Greco, Gaetano Valenza, Antonio Lanata, Enzo Pasquale Scilingo, and Luca Citi. 2015. cvxEDA: A convex optimization approach to electrodermal activity processing. IEEE Transactions on Biomedical Engineering 63, 4 (2015), 797–804.
  • Nils Y Hammerla, Reuben Kirkham, Peter Andras, and Thomas Ploetz. 2013. On preserving statistical characteristics of accelerometry data using their empirical cumulative distribution. In Proceedings of the 2013 international symposium on wearable computers. 65–68.
  • Javier Hernandez, Ivan Riobo, Agata Rozga, Gregory D Abowd, and Rosalind W Picard. 2014. Using electrodermal activity to recognize ease of engagement in children during social interactions. In Proceedings of the 2014 ACM international joint conference on pervasive and ubiquitous computing. 307–317.
  • Francisco Hernando-Gallego, David Luengo, and Antonio Artés-Rodríguez. 2017. Feature extraction of galvanic skin responses by nonnegative sparse deconvolution. IEEE journal of biomedical and health informatics 22, 5 (2017), 1385–1394.
  • Anne Horvers, Natasha Tombeng, Tibor Bosse, Ard W Lazonder, and Inge Molenaar. 2021. Detecting emotions through electrodermal activity in learning contexts: A systematic review. Sensors 21, 23 (2021), 7869.
  • BIOPAC Systems Inc. [n. d.]. EDA Introductory Guide. Retrieved May 25, 2023 from https://www.biopac.com/wp-content/uploads/EDA-Guide.pdf
  • Shamsi T. Iqbal, Xianjun Sam Zheng, and Brian P. Bailey. 2004. Task-Evoked Pupillary Response to Mental Workload in Human-Computer Interaction. In CHI ’04 Extended Abstracts on Human Factors in Computing Systems (Vienna, Austria) (CHI EA ’04). Association for Computing Machinery, New York, NY, USA, 1477–1480. https://doi.org/10.1145/985921.986094
  • Quintus R Jett and Jennifer M George. 2003. Work interrupted: A closer look at the role of interruptions in organizational life. Academy of management Review 28, 3 (2003), 494–507.
  • Jing Jin and Laura A Dabbish. 2009. Self-interruption on the computer: a typology of discretionary task interleaving. In Proceedings of the SIGCHI conference on human factors in computing systems. 1799–1808.
  • Harmanpreet Kaur, Alex C Williams, Daniel McDuff, Mary Czerwinski, Jaime Teevan, and Shamsi T Iqbal. 2020. Optimizing for happiness and productivity: Modeling opportune moments for transitions and breaks at work. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. 1–15.
  • Malia Kelsey, Murat Akcakaya, Ian R Kleckner, Richard Vincent Palumbo, Lisa Feldman Barrett, Karen S Quigley, and Matthew S Goodwin. 2018. Applications of sparse recovery and dictionary learning to enhance analysis of ambulatory electrodermal activity data. Biomedical Signal Processing and Control 40 (2018), 58–70.
  • Kyung Hwan Kim, Seok Won Bang, and Sang Ryong Kim. 2004. Emotion recognition system using short-term monitoring of physiological signals. Medical and biological engineering and computing 42 (2004), 419–427.
  • Ian R Kleckner, Rebecca M Jones, Oliver Wilder-Smith, Jolie B Wormwood, Murat Akcakaya, Karen S Quigley, Catherine Lord, and Matthew S Goodwin. 2017. Simple, transparent, and flexible automated quality assessment procedures for ambulatory electrodermal activity data. IEEE Transactions on Biomedical Engineering 65, 7 (2017), 1460–1467.
  • Matthew Lee. 2020. Detecting affective flow states of knowledge workers using physiological sensors. arXiv preprint arXiv:2006.10635 (2020).
  • Mindware Technologies Ltd.2023. "Mindware EDA". https://support.mindwaretech.com/training/guides/skin-conductance-eda-training-guide/#
  • Dominique Makowski, Tam Pham, Zen J Lau, Jan C Brammer, François Lespinasse, Hung Pham, Christopher Schölzel, and SH Annabel Chen. 2021. NeuroKit2: A Python toolbox for neurophysiological signal processing. Behavior research methods (2021), 1–8.
  • Gloria Mark, Shamsi T Iqbal, Mary Czerwinski, and Paul Johns. 2014. Bored mondays and focused afternoons: the rhythm of attention and online activity in the workplace. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. 3025–3034.
  • Luca Menghini, Evelyn Gianfranchi, Nicola Cellini, Elisabetta Patron, Mariaelena Tagliabue, and Michela Sarlo. 2019. Stressing the accuracy: Wrist-worn wearable sensor validation over different conditions. Psychophysiology 56, 11 (2019), e13441.
  • Nir Milstein and Ilanit Gordon. 2020. Validating measures of electrodermal activity and heart rate variability derived from the empatica E4 utilized in research settings that involve interactive dyadic states. Frontiers in Behavioral Neuroscience 14 (2020), 148.
  • Emily Claire Mitaro. 2019. Synchrony of Electrodermal Activity in a Socioeconomically and Racially Diverse Sample of Mothers and Preschoolers. Ph. D. Dissertation. University of Georgia.
  • Mohsen Nabian, Yu Yin, Jolie Wormwood, Karen S. Quigley, Lisa F. Barrett, and Sarah Ostadabbas. 2018. An Open-Source Feature Extraction Tool for the Analysis of Peripheral Physiological Data. IEEE Journal of Translational Engineering in Health and Medicine 6 (2018), 1–11. https://doi.org/10.1109/JTEHM.2018.2878000
  • Andrew F. H. Payne, Anne M. Schell, and Michael E. Dawson. 2016. Lapses in skin conductance responding across anatomical sites: Comparison of fingers, feet, forehead, and wrist. Psychophysiology 53, 7 (2016), 1084–1092. https://doi.org/10.1111/psyp.12643 arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/psyp.12643
  • Thomas Ploetz, Nils Hammerla, and Patrick Olivier. 2011. Feature Learning for Activity Recognition in Ubiquitous Computing.1729–1734. https://doi.org/10.5591/978-1-57735-516-8/IJCAI11-290
  • Thomas PlÖtz. 2021. Applying machine learning for sensor data analysis in interactive systems: Common pitfalls of pragmatic use and ways to avoid them. ACM Computing Surveys (CSUR) 54, 6 (2021), 1–25.
  • Ming-Zher Poh, Nicholas C Swenson, and Rosalind W Picard. 2010. A wearable sensor for unobtrusive, long-term assessment of electrodermal activity. IEEE transactions on Biomedical engineering 57, 5 (2010), 1243–1252.
  • Hugo F Posada-Quintero, John P Florian, Alvaro D Orjuela-Cañón, Tomas Aljama-Corrales, Sonia Charleston-Villalobos, and Ki H Chon. 2016. Power spectral density analysis of electrodermal activity for sympathetic function assessment. Annals of biomedical engineering 44 (2016), 3124–3135.
  • Hugo F Posada-Quintero, John P Florian, Alvaro D Orjuela-Cañón, and Ki H Chon. 2018. Electrodermal activity is sensitive to cognitive stress under water. Frontiers in physiology 8 (2018), 1128.
  • Izabela Rejer and Jarosław Jankowski. 2016. EEG patterns analysis in the process of recovery from interruptions. In Proceedings of the 9th International Conference on Computer Recognition Systems CORES 2015. Springer, 587–596.
  • Thought Technologies. 2023. "Flexcomp Infiniti". https://thoughttechnology.com/procomp5-infiniti-system-w-biograph-infiniti-software-t7525/
  • Yihan Wu and Wilma Koutstaal. 2020. Charting the contributions of cognitive flexibility to creativity: Self-guided transitions as a process-based index of creativity-related adaptivity. PloS one 15, 6 (2020), e0234473.
  • Lili Zhu, Petros Spachos, Pai Chet Ng, Yuanhao Yu, Yang Wang, Konstantinos Plataniotis, and Dimitrios Hatzinakos. 2023. Stress Detection through Wrist-based Electrodermal Activity Monitoring and Machine Learning. IEEE Journal of Biomedical and Health Informatics (2023).

CC-BY license image
This work is licensed under a Creative Commons Attribution International 4.0 License.

ISWC '23, October 08–12, 2023, Cancun, Quintana Roo, Mexico

© 2023 Copyright held by the owner/author(s).
ACM ISBN 979-8-4007-0199-3/23/10.
DOI: https://doi.org/10.1145/3594738.3611360