Electrodermal activity (EDA), commonly known as galvanic skin response, is a measure of sympathetic nervous system activity that is used to assess physiological arousal. Measurement of and methods for examining EDA in the laboratory setting have been well described [1], but there is less guidance for ambulatory EDA. The aim of this study was to describe and illustrate several analytic methods, specifically functional data analysis (FDA) [2] and local polynomial regression with autoregressive (AR) errors [3], neither of which have previously been applied to EDA data. We also addressed the challenges of collecting ambulatory EDA, denoising the data, and detecting meaningful features common across participants. We concluded with opportunities and limitations of the future use of ambulatory EDA data.

EDA has been measured in the laboratory using gel electrodes typically placed on the palm. Entire books [1] and publication standards (eg, Society for Psychophysiological Research Ad Hoc Committee on Electrodermal Measures [4]) have been written about the measurement and the use of EDA data collected in the laboratory. However, much less has been written about the measurement of EDA in ambulatory settings [5], which has recently become popular because of the availability of dry-electrode wrist-worn devices that can be used while participants go about their daily activities. These devices have dry electrodes that are placed on the ventral wrist as opposed to laboratory placement on the palm. Emotional effects on EDA can be more sensitively recorded with electrodes on the palm rather than the wrist, but obviously, electrodes on the palm would interfere with normal activities. However, dry electrodes used in ambulatory measurement have uncertain contact with the skin in comparison with gel electrodes used in laboratory-based measurements. Important issues in the measurement of ambulatory EDA are the influence of temperature (both ambient and skin) and physical activity [4]. Several wrist-worn devices have been developed for the ambulatory collection of EDA as well as temperature and three-axis accelerometry [6,7] (eg, Affectiva Q sensor) and, more recently, blood volume pulse (eg, Empatica E4). This study used the iCalm device [8], which measures temperature and accelerometry as well as EDA (the Methods section provides a detailed description about this).

All these devices record the overall skin conductance (SC), which can be decomposed into what is referred to as tonic and phasic activity components [9-11]. Tonic activity varies slowly and is also referred to as skin conductance level (SCL). Individuals have different levels of tonic activity, that is, some individuals have a higher level of tonic activity and others have a lower level of tonic activity. In contrast, phasic activity varies rapidly in response to stimuli such as stress and is also referred to as the skin conductance response (SCR). The SCR amplitude is considered to be a measure of sympathetic activity. SCR is characterized by a rapid incline to a peak and then a slower decline back to the individual’s SCL [1]. Typically, the goal is to detect these peaks and compute their features (eg, amplitude), which can serve as predictors or outcomes in statistical models. Although the features can be detected by visualizing the data and coding the peaks, the amount of data that are typically collected in ambulatory research studies makes this time consuming and impractical. Next, we introduce the study from which the data were obtained before returning to approaches for analyzing ambulatory EDA data.

Many adolescent mothers (AMs) have challenges effectively regulating emotion, which interferes with their use of positive parenting skills and places them at risk of maltreating their children. Data for this study were collected from participants in a pilot, group-randomized controlled trial of the Power Source Parenting (PSP) intervention for high-risk AMs [12,13]. The AMs were homeless and lived with their children in transitional living programs (TLPs) in a Northeast state. Each participant received a baseline and 2 postintervention follow-up assessments (at 3 and 6 months). Each assessment lasted between 1.5 and 2 hours, and the AMs wore the iCalm biosensor band that measured EDA throughout each of the 3 assessments, which included 2 potentially stressful tasks: (1) a timed Stroop task and (2) a 10-min videotaped mother-child interaction. We examined EDA at each assessment for the Stroop and video interaction tasks. We chose these 2 tasks because we expected that they induce different types of stressors (eg, Stroop task is a cognitive stressor) and they have differing degrees of structured activity. The Stroop task was more structured than the videotaped interaction task, and it more closely resembles the measurement of EDA in the laboratory (ie, AMs complete the computerized task seated alone without her child present). In contrast, the semistructured mother-child videotaped interaction does not have the constraints of carefully controlled laboratory settings but provides more ecologically valid information about maternal and child behaviors concurrent with measurement of AMs’ EDA that can be examined across participants.

In our study, different AMs had quite different features in their EDA profiles (ie, the EDA level over the time span of a single task), making it difficult to compare them directly. Some of the differences are noise, such as those from movement artifacts. Despite innovations in devices, ambulatory assessment of EDA is inherently noisier than laboratory assessment because of these movement artifacts and the dry electrodes. Thus, the first step in processing EDA data is to identify these artifacts so that these data can be discarded. The step typically involves applying a signal processing filter, such as a Hanning filter with a 1-second window [14], a zero-phase order 10 low-frequency Butterworth filter [5], or a finite impulse response (FIR) low-pass filter [15]. The application of a signal processing filter is not mandatory, as the next step, identifying valid versus invalid data, can be applied directly to the raw, unfiltered data [15]. Taylor et al [15] developed an algorithm using support vector machines (SVMs) to identify valid versus invalid data and applied it to both the raw SC data and SC data filtered using an FIR low-pass filter. SVMs have a disadvantage in that it is not transparent what rules are being implemented to classify data as valid versus invalid. In contrast, a rule-based algorithm was proposed [16] for classifying data as valid or not. For example, values of less than 0.05 μS could be excluded [5]. A disadvantage of this type of approach is that it is less flexible; however, it does not have the reproducibility problems that the Taylor et al [15] approach may have [16]. Usually, a combination of a signal processing filter and either a rule-based approach or a machine learning approach is used to identify and discard invalid data [5,15]. Given the valid data, the next step is to separate the tonic and phasic components.

Many of the previous methods for separating the components are based on EDA data collected in the laboratory in which the participant is exposed to a stimulus, such as a loud noise, at regular intervals (eg, interstimulus interval of 8 seconds). These methods include the sigmoid exponential model [11] and negative deconvolution [9]. Although these methods were developed for data collected in the laboratory, the negative deconvolution method has been applied to ambulatory data [14]. There is a question as to whether it is important to separate these components in ambulatory data because the stimuli are typically more gradual (typically referred to as nonspecific SC responses [1]) as opposed to the abrupt startle response stimuli typically used in laboratory research. Hernandez et al [14] analyzed the data without decomposing the tonic and phasic components and by decomposing them using nonnegative deconvolution and found that the decomposition provided more discriminative information. Thus, they recommended decomposing the EDA. Regardless of whether the EDA is decomposed or not, the next step is to identify features of the data, typically peaks, and compute the characteristics of these features.

One automated approach to detecting peaks and constructing features is available on an interactive website at the MIT Media Lab [15] and uses SVMs, although the focus was primarily on identifying valid versus invalid data. Kleckner et al [16] also did not focus on detecting peaks but only on identifying valid versus invalid data. Thus, there is very little guidance on detecting peaks once data have been identified as valid. We initially used the Taylor et al [15] approach and discovered that the approach did not identify many of the peaks and marked much of the data as invalid. The results of applying this approach are presented in Figure 1 for one AM as an example and are what motivated us to try to find a better approach to detecting peaks in the ambulatory setting.

In general, our approach was to (1) use the raw SC data or SC data filtered with an FIR low-pass filter, (2) identify valid versus invalid data using rules established in the literature, (3) decompose the signal into its tonic and phasic components using negative deconvolution, (4) apply a spline-based smoothing technique (eg, B-spline regression or local polynomial regression with AR errors) to the EDA profile, and (5) compute various features of the peaks identified from the smoothed curve. The latter 2 steps, particularly the spline-based approaches, are the unique contributions of this study. Comparisons between different groups of individuals can then be performed on the peak features, if desired.

The iCalm applies a constant voltage to the skin through 2 silver/silver chloride metal electrodes and measures the resulting current [8,13]. A transconductance amplifier is used to convert the current to a voltage, which is sampled at a rate of 4 Hz using a 12-bit analog-to-digital converter. Three-axis accelerometry and temperature were sampled at the same rate. The iCalm has been validated against laboratory measures of EDA [6]. The iCalm connects via Bluetooth to Android-based mobile phones.

The data were obtained from the aforementioned PSP intervention study of 71 mothers. The EDA values for some mothers were very low, which was primarily because of a lack of contact between the skin and the electrodes. The general guidance in the literature is to discard data in which the EDA is approximately 0 [16], as it indicates that the electrodes were not in good contact with the skin. Details regarding these invalid data are given in section Data Preprocessing. Finally, there were 5 AMs who did not return at any of the follow-up assessments and several AMs who had data at one but not both of the follow-up assessment periods; thus, we used 6-month follow-up data if an AM did not have data at the 3-month follow-up. The final sample size numbers are given following discussion of the identification of valid versus invalid data.

The AMs were interviewed at the TLP in a private room by trained interviewers. The interviews included a survey questionnaire examining demographic and background characteristics; questions regarding mental health, parenting, and risk-taking behaviors; as well as activities described in the Measures section. AMs wore the iCalm band on their right wrist throughout the duration of each interview. Following the interview, mothers in the intervention TLPs received the PSP intervention, whereas the control TLPs received standard care [12].

Each assessment included the computer-delivered survey, a variety of computerized tasks including the Stroop task, and the 10-min videotaped interaction of mothers with their child. In this analysis, we focused on the Stroop task and the videotaped interaction task. The Stroop task involved reading the written color names independent of the color name that is written as quickly as possible in 1 min. For the videotaped interaction task, the mothers were asked to play with their child as they usually do for the first 5 min, which we will refer to as free play. For the second 5 min of the interaction task, the mothers were asked to teach their child a concept that was slightly above their developmental level. This latter task, which we will refer to as teaching, was intended to be slightly frustrating for the mother and the child. In addition, a physical task (going up and down stairs for approximately 5 min) was only performed at the baseline assessment and only for the intervention group AMs. We examined these data as a comparison task in which there is a lot of movement.

The data analysis for analyzing and comparing the EDA profiles entails several steps described earlier, including preprocessing, smoothing, identifying peaks, computing features of the peaks, and, finally, statistical modeling. In this section, we present the details of these steps with regard to our data. The R code for implementing the analysis is provided in Multimedia Appendix 1. To date, software for identifying invalid data, decomposing SC, detecting peaks, and constructing features has been written in either MATLAB or Python. To our knowledge, there are no R packages for handling these tasks.

Next, we present the results of the local linear regression with AR errors. Figure 3 shows the smoothed function overlaying the normalized EDA values for the baseline and 3-month follow-up assessments for the teaching task for one control group AM. Figure 4 presents the smoothed function overlaying the normalized EDA values for baseline and 3-month follow-up assessments for the free-play task for the same AM, as presented in Figure 3. In the case of this AM, the smoothed function does not fit the actual peaks in the data very well. Specifically, the smoothed peaks are dampened in comparison with the actual peaks, which creates a problem for identifying peaks because they are defined by their increase/decrease. Even if peaks are identified, some features of the peaks (eg, amplitude) would be dampened in comparison with the actual data. We also tried using an FIR low-pass filter before smoothing with local linear regression for AR errors. However, this did not substantially improve the fit of the estimated functions to the actual data. Figure 5 shows the smoothed function overlaying the normalized EDA values after applying the FIR low-pass filter for the baseline and 3-month follow-up assessments for the teaching task for the same AM, as presented in Figure 3. The free-play task is presented in Figure 6. It is evident that the FIR low-pass filter did not change the estimated EDA profiles substantially from the raw SC data. We next attempted smoothing using B-splines to estimate the functional EDA profiles.

In implementing the smoothing via B-splines, we used 13 basis functions. We plotted the smoothed versus normalized EDA for each AM, for each task, and for each assessment. The plots for the same AM illustrated in Figures 3 and 4 are shown in Figure 7, left column. The figure illustrates a potential problem with the B-spline approach, which relies on a cubic polynomial. To fit the first peak, the cubic function, by definition, must first increase and then decrease before increasing again. In so doing, the function imposes a false peak just before the actual first peak, and therefore, the function does not fit the actual data well (Figure 7). A different polynomial could be chosen for fitting the B-splines, but it would also impose a particular functional form that may not fit the actual data well. Figure 8 shows the EDA profiles for the teaching task for the same AM (PID 102), as shown in Figure 7. In comparison with the local linear regression approach illustrated in Figure 3, the B-spline approach illustrated in Figure 8 appears to fit at least as well for PID 102. In addition, the B-spline approach fits some individuals better than the local linear regression (not shown). For example, EDA for the free-play task for one intervention group AM is presented in Figure 9, and EDA for the teaching task for the same AM is presented in Figure 10. For this AM, the B-spline approach seemed to work well. We will discuss the 2 smoothing approaches further in the Discussion section.

Next, we identified peaks and computed various features of the peaks from the estimated EDA profiles. Figure 7 (right column) illustrates the identified peaks and presents the values for the peak features. The red dashed vertical lines demarcate the first peak, the black dashed vertical lines demarcate the highest peak, and the blue dashed vertical lines demarcate all other identified peaks. The time to first peak and the highest peak (both in seconds), the normalized EDA value at the highest peak (ie, amplitude), the drop from the highest peak (ie, reactivity), and the number of peaks are all presented in Figure 7, right column for one AM. The peaks were identified, and the features were computed for all AMs. These may be saved in a file for statistical analysis. Our sample size was ultimately too small for statistical significance testing between the intervention and control groups (ie, between-subjects), but we present descriptive statistics of the peak features in Table 1.

As mentioned previously, a lot of data were lost because the electrodes did not have a good connection with the skin, which resulted in EDA values that were approximately 0. Thus, the first lesson is to make sure that there is a good connection between the dry electrodes on the device and the skin. This means that the electrodes must be touching the skin on the ventral side of the wrist, which means that the device must be fairly tight on the wrist. People often do not wear watches, bands, or bracelets as tight as is necessary for the EDA device. Therefore, we suggest real-time monitoring of incoming EDA data so that if all the values are approximately 0, the participant could be contacted to inform them to tighten the device. Note that simply detecting that data are not being recorded in real time is not enough; there must be real-time detection of data that is being recorded but is essentially 0. Some of our EDA data were not usable because the band lost contact with the skin during some tasks, particularly early in the interview. By loose connection, we mean that the electrodes on the band did not have a good connection with the skin. We do not mean that the band was not connected in terms of recording the EDA. If the band was not recording, then EDA was missing at that particular measurement. This did happen sometimes but not enough to present a problem, particularly because we were able to smooth over it using the proposed smoothing approaches. It should be noted that the loose connection problem is not limited to the particular device, the iCalm, that we used. Empatica E4 also has this problem, as we have used it in a pilot study [24] and lost data because of a loose connection. Other studies using wrist-based ambulatory measures of EDA with the Affectiva Q Sensor have reported greater success [25,26] in both controlled and uncontrolled settings. However, other studies using the Affectiva Q sensor have reported data lost to electrode connection problems (eg, 31% [14] and 17% [22]). A total of 22.5% of data were lost because of electrode connection problems for the free-play task and 24% for the teaching task.

Owing to the loose connection problems during data collection, our sample size was substantially reduced, which limits our ability to draw conclusions about the between-subject effects of the intervention on the features constructed from the EDA data. Nevertheless, we think it is important to draw attention to data collection challenges as the measurement of ambulatory EDA has become more popular. In addition, researchers currently collecting EDA data have limited guidance on how to preprocess the data before statistical analysis.

A second challenge in EDA data collection is motion artifacts. These are identifiable as sharp spikes or drops in EDA data. One concern with the free-play task is that there were many more potential motion artifacts than there were in the teaching task because the AMs may have been moving around more during the free-play task. Another issue that may sometimes occur is that the EDA during the current task could be affected by the EDA response to the previous task. In our study, we tried to minimize this by having recovery periods between tasks, but, of course, there is no guarantee that the recovery periods were long enough. An individual’s EDA does not always return to the previous baseline, so it can be difficult to assess if the recovery period is long enough.

In summary, in light of the increase in wearables for the continuous, ambulatory collection of EDA and associated challenges collecting data outside of the laboratory, analytic methods for computing various features of EDA are needed to help meet these challenges. We described 2 methods for smoothing and summarizing EDA data using FDA and local polynomial regression with AR errors. As the technology of wearable devices continues to advance, future research opportunities abound for improvement in device design that may improve the electrodes and their connection with the skin and facilitate in the development of improved methods for smoothing the data, identifying peaks, and constructing features.