Unsupervised Electrodermal Data Analysis Comparison between Biopac

and Empatica E4 Data Collection Platforms

Kassy Raymond and Andrew Hamilton-Wright

School of Computer Science, University of Guelph, Guelph, Ontario, Canada

Keywords:

Clustering, Quality Metrics, Biosignal Analysis, Unsupervised Machine-learning, Data Analytics.

Abstract:

Unsupervised learning algorithms are valuable for exploring a variety of data domains. In this paper we

compare the efﬁcacy of the k-means and DBSCAN algorithms in the context of discerning structure in elec-

trodermal data obtained using two different collection modalities for simultaneously collected data: the “gold

standard” Biopac data platform and the wearable Empatica E4. Insights into the structure of the data from

each system are provided, as is an analysis of the performance of each clustering algorithm at identifying

interesting structure within the data.

1 INTRODUCTION

Electrodermal activity (EDA) is a psychophysiologi-

cal index that is used as a measure of arousal in the

sympathetic nervous system and thus as an indicator

of whether an individual is eliciting a physiological

response to stress. In laboratory settings, EDA can

be measured using electrodes adhered to the skin us-

ing isotonic paste, such as those offered by Biopac

(Biopac, 2019) or through a wrist wearable such as

the Empatica E4 (Empatica, 2020). While adhered

electrodes are considered the “gold standard” acqui-

sition device, it is unknown whether similar insights

can be drawn from signals acquired from the E4 using

an unsupervised machine learning approach.

We compare the data available through these two

collection protocols as analyzed through a variety of

sample window lengths, and using two different clus-

tering techniques: k-means and Density-based Spatial

Clustering of Applications with Noise, or DBSCAN

(Ester et al., 1996; Sander et al., 1998). Using these

techniques, we evaluate the ability to discern structure

within a data set collected using the Biopac and E4

modalities at the University of Guelph by Dr. Kristal

Thomassin’s Child Emotion and Mental Health Lab.

When this data was collected, the labels deﬁn-

ing the activity regions were lost, resulting in a prob-

lem of reassigning labels based purely on the struc-

ture found using unsupervised techniques. This work

explores the success of determining structure in this

challenging scenario, and through heatmaps, show-

cases the discovered data structure.

https://orcid.org/0000-0002-7459-656X

2 BACKGROUND

Electrodermal Activity (EDA) is a measure of the ac-

tivation of sweat glands through changes in conduc-

tivity of the skin due to variations in sweat secretion,

typically obtained using Ag/AgCl (silver/silver chlo-

ride) electrodes (Cecchi et al., 2020; Posada-Quintero

and Chon, 2020; Lowry, 1977).

Data for this study were simultaneously obtained

using two different collection platforms. The ﬁrst

is the research industry standard biosignal acquisi-

tion device, Biopac, which uses electrodes attached to

belt-mounted transceivers in order to collect its data.

The second is the Empatica E4, which is a research

grade wrist mounted wearable device, which is both

cheaper and simpler to attach. We will refer to this

device simply as the E4 in this work.

2.1 EDA Data

EDA signals consist of two main components: the

tonic component termed SCL for “Skin Conductance

Level” (SCL) and the phasic component – SCR for

“Skin Conductance Response” (Boucsein, 2012).

The SCL, or tonic component of the signal, is

the low frequency component of the signal, identi-

ﬁed by the “baseline,” or the general shape of the sig-

nal (Boucsein, 2012; Cho et al., 2017). SCL activ-

ity is generally independent of environmental stimuli;

while SCL activity can slowly increase during periods

of emotional arousal, homeostasis and overall mois-

ture content of the skin are the main drivers of SCL

activity (Boucsein, 2012).

The SCR is associated with the activity of the

Raymond, K. and Hamilton-Wright, A.

Unsupervised Electrodermal Data Analysis Comparison between Biopac and Empatica E4 Data Collection Platforms.

DOI: 10.5220/0011271800003269

In Proceedings of the 11th International Conference on Data Science, Technology and Applications (DATA 2022), pages 345-352

ISBN: 978-989-758-583-8; ISSN: 2184-285X

345

eccrine sweat glands located in the dermis of the

skin which are stimulated by the sympathetic ner-

vous system via cholinergic neural pathways during

stress or activity (Schmelz et al., 1998; Millington

and Graham-Brown, 2010; Charmandari et al., 2005;

Benedek and Kaernbach, 2010; Chen et al., 2020).

The SCL and SCR traces can be produced us-

ing deconvolution via the cvxEDA algorithm (Greco

et al., 2016), which is implemented and made publicly

available within Neurokit2, an open source Python

package (Makowski et al., 2020). Using these tools in

combination with the pyphysio toolkit (Bizzego et al.,

2019), it is relatively simple to separate EDA data into

source traces and accompanying measures of the am-

plitude and rate of observed pulses within the signal.

It is therefore of interest to characterize the changes

in measured skin response, both SCL and SCR, in or-

der to obtain insight into the level of stress (Boucsein,

2012). We therefore are interested in analytical tools

that examine the structure of such data.

2.2 Analytical Tools

In order to assess the free-form, unlabelled data col-

lected by Dr. Thomassin’s lab, we turn to unsuper-

vised learning (clustering) tools. Our tools of inter-

est here are the well-understood k-means clustering

(MacQueen, 1965), and Density-Based Spatial Clus-

tering of Applications with Noise, or DBSCAN (Ester

et al., 1996; Sander et al., 1998).

The k-means algorithm simply deﬁnes a cluster by

the location of its centre, and the data is assumed to

form a “Gaussian ball” radiating symmetrically out

from this point in all directions. The k-means algo-

rithm produces a centroidal Voronoi Tesselation (Du

et al., 1999) in which cluster region boundaries are de-

ﬁned by straight lines or planes of intersection where

the two Gaussian “balls” meet, much like two soap

bubbles that get joined together.

The DBSCAN algorithm, on the other hand, aims

to identify areas of high density that are separated by

areas of low density in a feature space, without mak-

ing any a priori assumptions about the shape of the

cluster or its relative position with respect to other

clusters. DBSCAN is especially effective for feature

spaces with arbitrary and/or asymmetrical shapes as

its deﬁnition of the cluster region follows the topol-

ogy of the densest portions of the cluster body.

Because DBSCAN is concerned with density, it

is imperative to be able to quantify and deﬁne the pa-

rameters related to density, such as a “dense area,” and

“density at a point” (Ester et al., 1996). “Density at a

point” is deﬁned as the number of points within a cir-

cle of radius ε from a given point, P. A “dense area”

is deﬁned by a minimum number of points (MinPts),

where each P in a cluster must contain MinPts within

its radius ε. MinPts and ε are the only input parame-

ters needed for DBSCAN. Unlike k-means, for which

the number of clusters is a parameter that must be

determined in advance, DBSCAN selects the optimal

number of clusters based on these values.

3 METHODS

EDA data was collected from 43 healthy undergradu-

ate students at the University of Guelph who voluntar-

ily participated in the study. The mean age of the par-

ticipants was 18.87 (SD = 0.88). 15 of the participants

were male, 27 were female and one was non-binary.

The study protocol, which was authorized by the

University of Guelph Research Ethics Board, con-

sisted of: 3 minutes sitting quietly; the Trier Social

Stress Test (Kirschbaum et al., 1993) to elicit stress; 3

minutes to prepare for a mock job interview, followed

by 5 minutes for the interview itself; a 5 minute arith-

metic task involving counting down from 1022 in in-

crements of 13 as quickly as possible; 1 minute for

completion of a PANAS survey (Watson et al., 1988);

a 5 minute recovery period.

This data provides a mixture of high and lower

stress responses obtained from participants wearing

the E4 on their dominant wrist and using a standard

placement for Biopac electrodes on the non-dominant

hand, chest and stomach. Participants placed the elec-

trodes themselves for reasons of privacy.

The data for nine participants was removed from

the study due to recording problems, resulting in 34

collections available for analysis. The remaining data

was then separated into SCL and SCR traces using the

tools described above.

3.1 Windowing

The effect of the size of a windowing protocol on the

ability of the clustering algorithms to extract infor-

mation was studied. In order to evaluate the effect

of window size, each signal for each participant was

separated into overlapping windows with a window

length (WL) of 3, 5, 10, 30, 45, 60, 90 and 120 sec-

onds. The step in the window varied depending on

the length of the window; the overlap was half of the

length of the window in all cases.

Within each window, the following features were

obtained using pyphysio from the phasic EDA sig-

nal:mean, standard deviation, range, peak mean, peak

maximum, peak minimum, number of peaks, dura-

tion mean, slope mean, and area under curve (AUC).

DATA 2022 - 11th International Conference on Data Science, Technology and Applications

346

The “phasic” values were calculated by ﬁrst identify-

ing each peak (i.e.; each phase) in the signal window,

and then calculating the value over this set of identi-

ﬁed peaks. Area under the curve is the integral of the

signal, calculated by the sum of the absolute values of

each sample (i.e.; the sum of the “rectiﬁed signal”).

3.2 Input Data Set Construction

Data was considered in the following groupings:

DS1a: for each window length, data from all partici-

pants by device: this gives 8 window lengths each

for Biopac and E4

DS1b: the same data, but using pooled analysis and

data display: all data combined (across window

length and acquisition devices) — this allow plot-

ting on common axes

DS2a: as in DS1a, however with participant 15 re-

moved (see discussion below)

DS2b: as in DS1b with participant 15 removed

PCA was conducted on all datasets described above.

To determine the optimal number of components, the

explained variance in each of the principal compo-

nents was calculated, and the number of components

required to explain 99% of the variance was used.

PCA is also used to obtain a two-dimensional pro-

jection used in plotting below.

3.3 Clustering

To determine the suitability of clustering algorithms,

DS1a data were clustered with DBSCAN and k-

means for each collection modality and window

length. An “elbow” plot was used to determine the

optimal k for each data set, selecting the point of max-

imum curvature. Silhouette plots were also consid-

ered, however instabilities in the clusters created by

k-means made these unusable—an early tell regard-

ing the limitations of k-means.

To calculate ε, an “elbow” was again used as is

common practice (Ester et al., 1996). MinPts was cal-

culated using the heuristic suggested by (Sander et al.,

1998) where MinPts equals the dimensions of the

dataset multiplied by two. This approach was taken

in attempts to remain consistent across the datasets

for each device and window length.

The cluster assignments were evaluated by plot-

ting the assignments for each datasets with a scatter

plot where the axes represented the principal compo-

nents calculated with DS1b (i.e.; using common co-

ordinates). The plots were visually assessed to deter-

mine which algorithm was most suitable for the data;

this included comparing the structure of the clusters,

identifying logical clusters and discussing the ability

for the algorithms to logically cluster points that ap-

peared sparsely distributed in the feature space (i.e.,

identify noise points).

To determine which participants were expressed

in each cluster, the number of points in each cluster

corresponding to each participant were tallied. This

gives an indication of whether the information con-

tent across all signals from all participants were be-

ing clustered or whether the signals of an entire par-

ticipants were being clustered. While it is expected

that stressful and baseline events should exhibit dif-

ferent patterns in the signals, there could be variation

in whole signals between participants. For all window

lengths and each device, the cluster assignments for

each participant were plotted in a time-series heatmap

to visually interpret which portions of the signal cor-

responded to each cluster.

4 RESULTS AND ANALYSIS

Fig. 1 displays scatter plots of the cluster category la-

bels as shown plotted across the two most principal

axes calculated on a PCA decomposition of data set

DS2b. As both collection modalities are grouped, this

provides common axes for each window.

Immediately apparent in Fig. 1 is that the axes are

stable across windows, as shown by the common lay-

out of the scatter plots. The differing label colours

used are simply a function of which label was as-

signed ﬁrst in each clustering algorithm, so colour is

independent between plots.

The straight-line divisions shown in the k-means

plots that cut across clouds of data are a limitation of

the k-means algorithm, and the ability of DBSCAN

to extract features such as the line forming a lead-

ing edge of the leftmost spray of points in the Biopac

data is a show of the relative ability of this high-

performing algorithm to extract meaningful structure

based on density.

The similarity of all of the plots within a given col-

lection modality (Biopac, E4) but the distinct struc-

ture between modalities indicates that there are signif-

icant differences between these collection framework

data distributions.

Fig. 2 shows select heatmaps displaying the asso-

ciation of a cluster category in coloured patches ar-

ranged over time (x-axis) for each participant (rows

down the y-axis). Plots are then repeated in columns

for each data modality showing the results for rep-

resentative window length (WL) of {3, 10, 45, 90}.

(Other lengths are similar, and omitted due to space

constraints.)

While these images are tiny, this is only in part

Unsupervised Electrodermal Data Analysis Comparison between Biopac and Empatica E4 Data Collection Platforms

347

Biopac

k-means

DBSCAN

k-means

DBSCAN

WL=3 WL=5 WL=10 WL=30

Figure 1: Scatter plot data for dataset DS2b.

because of the space limitations of the paper, and

is intended to present the user with a “small multi-

ples” presentation of the important data as champi-

oned in Edward Tufte’s famous book (Tufte, 1991,

pp. 67). Using the small multiples technique, com-

monalities among the repeated visual representations

can be comprehended at once, and the signature dif-

ferences between the similar representations become

easily apparent.

Regardless of the small size of the graphic ele-

ments in Fig. 2, the presence of an outlying cluster-

ing categorization is readily visible in the ﬁrst row

(Biopac data) as a horizontal bar roughly half-way

through the heatmaps that appears as a strikingly dif-

ferently coloured band. This band is the data belong-

ing to participant 15, and it corresponds to the ex-

tremely short band in the E4 data in the second row.

Also of note is participant 12, which appears as a

similarly distinct band in the E4 data—this band is

most notable in WL=10 and WL=90 where it is lighter

coloured than the label chosen for the bulk of the data.

As noted above, colour is related only to the la-

bel within a plot, and colours between plots can be

exchanged without implying anything.

The ability to discern this within this small-

multiples setup shows the strength of this heatmap

based strategy as a means of identifying overall struc-

ture.

The anomalous data associated with problematic

participants 12 and 15 was conﬁrmed in analysis of

PCA based scatter-plotting of DS1b data of Fig. 1

where the signal from participant 15 forms its own

cluster. In Fig. 1 this is the oval cloud shown on the

right of each Biopac plot, which is markedly distinct

from the fan shaped spray describing the data of the

other participants.

The presence of these extreme outlying classes

obfuscates the structure of the bulk of the data, how-

ever, so we continue with participant 15 and 12 omit-

ted. Upon re-clustering the data and replotting, we

obtain the plots shown in Fig. 3.

The heatmaps in both Figs. 2 and 3 show vertical

DATA 2022 - 11th International Conference on Data Science, Technology and Applications

348

WL Biopac E4

Figure 2: Heatmap data for dataset DS1a (all participants).

banding through the middle of the maps. This is par-

ticularly visible in the E4 data at WL=5 and 10, and

remains visible up through higher window lengths,

however with less deﬁnition. These bands correspond

to the periods in the signal at which the high-stress

activities are present, and indicate that, especially in

the E4 data, these stress measures are structurally dif-

ferent than the non-stress periods.

Note that this vertical banding is not as apparent

in the Biopac data, though it is striking in the E4 data.

4.1 Choice of Clustering Algorithm

Clustering with DBSCAN results in logical clusters

while k-means ﬁnds multiple clusters in a single logi-

cal cluster as indicated in Fig. 1 window length 45.

DBSCAN was able to recognize outlying points in

the “right hand cloud”. k-means appears to have sim-

ply “sliced-up” the range in values; it is typical for

k-means to ﬁnd divisions in data values simply be-

cause that is how the algorithm was designed. This

is trend is apparent in all Biopac scatter plots shown

in Fig. 1; notably in window lengths 30, 45, 60, 90

and 120). The behaviour of k-means isn’t as obvi-

ous in E4 due to the dense structure of the scatter plot

WL Biopac E4

Figure 3: Heatmap data for dataset DS2a (participants

12, 15 excluded).

clusters. However, the limitations of k-means are re-

vealed in plots for window lengths 60, 90 and 120.

Again, compared to the clusters identiﬁed by DB-

SCAN, k-means provides a less logical cluster struc-

ture. Further, across all window lengths and with E4

and Biopac, DBSCAN successfully identiﬁes noise

points, which are “ﬂoating” points that are separate

from the main clusters identiﬁed in the space, while k-

means groups these points into pre-existing clusters.

k-means is limited due to the “Gaussian ball” as-

sumption of spherical symmetry. The straight-line di-

visions visible in some of the k-means plots are also a

direct result of the Gaussian ball. Viewed in Fig. 1

these are what cause the straight lines dividing re-

gions, as are seen in where the orange and brown

points are separated by what appears to be a straight

line. The features calculated from this data are subject

to different degrees of variability in noise or outliers.

For example, the number of peaks and the phasic stan-

dard deviation have no a priori reason to vary in the

same way due to noise, which is an assumption in the

Gaussian ball. Determining the k input value for the

k-means algorithm also highlighted the difﬁculty of

selecting input variables for unsupervised learning al-

Unsupervised Electrodermal Data Analysis Comparison between Biopac and Empatica E4 Data Collection Platforms

349

gorithms.

Visual assessment of the DBSCAN clusters re-

sulted in a logical clustering structure and the ability

to isolate noise, deeming DBSCAN a suitable algo-

rithm for the proceeding experiments conducted. In

spite of the continued popularity of the simpler k-

means algorithm, we recommend DBSCAN for its su-

perior handling of the complexity of the electrodermal

data examined here.

4.2 Structure Revealed

It is through the time series heatmaps and graphs of

cluster assignments that the signals between each ac-

quisition device are compared and that structure in the

data are revealed. Larger window lengths (60, 90 and

120) provide a more concrete basis for understanding

signal-based analyses such as the separation of base-

line and stressful tasks.

When the input data from Biopac and E4 were

clustered together with DBSCAN, the majority of the

data from each acquisition device was clustered to-

gether. This indicates that the regions of the signals

acquired from both Biopac and E4, across all win-

dow lengths, exhibit enough similarly for DBSCAN

to cluster these regions together. Biopac data does

form unique clusters that are not populated any E4

data points when clustered with feature data calcu-

lated with window lengths of 3, 5, 10, 30 and 60.

Although differences in electrodermal activity signals

may exist between Biopac and E4, in machine learn-

ing experiments where the interest is focused around

class identiﬁcation (such as identifying stressful and

non-stressful tasks), even with differing signals, each

may be able to identify stressful and baseline tasks

just perhaps not in the same ways.

The structures of the cluster scatter plots are visi-

bly different, indicating a differing structure of data

between E4 and Biopac (Fig. 2). However, when

problematic participants were dropped, the structure

of Biopac clusters more closely resembled that of E4

(Fig. 3).

Notably, the absence of the cluster visible on the

right side of the Biopac scatter plots of Fig. 1 indicates

that this cluster was largely populated by the data of

the signals of these participants, and the data is of dif-

ferent structure than the remainder of the signals in

the space. The absence of this cluster in E4 in general

indicates that signals identiﬁed in problematic partici-

pants differs from that of Biopac. Identiﬁcation of this

cluster through this data analysis is therefore valuable

as a means of improving data quality.

The presence of regular and identiﬁable blocking

of cluster assignments across the signal of each par-

ticipant when clustered with window lengths 60, 90

and 120, as displayed in time-series heatmaps, im-

plies support using clustering techniques as a way of

identifying structure in electrodermal data, even in the

absence of labels.

Clusters -1 and 1 rarely occur during the begin-

ning the signal. Following this period of transition to

cluster -1 is identiﬁed which appears in blocks, with

deviations back to Cluster 1 and 0. The periodicity is

not consistent across all participants but can be clearly

identiﬁed in participants 2, 5, 8, 10, 16, 17, 18, 23,

28, 29, 31 and 37. In participants displaying the pe-

riodic transitions, the noise cluster is often the cluster

that appears in the middle section of the signal. This

portion is where the stress period is expected. While

DBSCAN identiﬁes this portion as noise, the portion

of the signal may not be noise in the technical sense.

Variations in the degree of stress and physiological

response to stress are expected so these differences

may be identiﬁed as “noise” by the algorithm rather

than having unique, dense clusters for stressful events

across all participants.

In spite of these sources of vagueness, the pres-

ence of these regions of signal on the plots do sup-

port the idea that there are observable differences in

the clustering assignment due to stress and non-stress.

Further, the fact that there are similar banding struc-

tures in both Biopac and in E4 data indicate that sim-

ilar structure is present in both experimental modali-

ties, though not manifest in the same way. If anything,

the “banding” effect present in the E4 data presented

in both Figs. 2 and 3 is more visually discernible than

in the Biopac data of the same ﬁgures. Contrasting

the plots in Figure 4.6 with their E4 counterparts in

Figure 4.8 shows a clear on/off/on progression in E4

that is evident only as a lead-in period in one class fol-

lowed by relatively unstructured data after the lead-in

point.

This is particular interesting in light of the fact

that Biopac is the gold-standard platform for this type

of data collection. The presence of this more visible

structure using the E4 sensor implies that this attrac-

tive wearable data collection platform contains data

relevant to the stress/non-stress discernment problem.

4.3 Window Length

At a higher window length (WL=60, 90 or 120 sec-

onds), the results suggest that it is possible that unsu-

pervised learning, speciﬁcally DBSCAN, may be able

to separate baseline and stressful tasks in spite of the

lack of labels in the source data.

As DBSCAN provides superior clustering per-

formance for this problem, as noted above, we will

DATA 2022 - 11th International Conference on Data Science, Technology and Applications

350

only consider DBSCAN based windowing. Window

length is oftentimes overlooked when designing ex-

periments, however, selecting an appropriate window

length to capture the type of information in a sig-

nal impacts the overall result of the ability to iden-

tify structure, whether it be noise, or stressful versus

baseline periods in the signal.

The cluster structure changes as the window

length is adjusted in both Biopac and E4 data (Figs. 2

and 3). The clusters become more dense as the win-

dow length increases, with fewer noise points pop-

ulating the outer regions of the plots. The change

in dispersion in the layout of the points among the

various window lengths is most pronounced in the

Biopac plots. This change in structure could be due

to the fact that the sample size decreases as the win-

dow length increases. However, the absence of the

green cluster in Biopac plots in window lengths 10

and above indicates that the feature vectors carry dif-

ferent information content at higher window lengths

and that the data from these points have been incorpo-

rated into the longer sequences. Similarly, the struc-

ture of E4 clusters at a window length of 30 and above

is much different than at the lower window lengths;

the smaller clusters are clustered into one. Again, this

indicates that the structure of the data is contingent on

the window length selected. It is quite likely that win-

dow lengths of 3, 5 and 10 are too small to capture

the physiological information provided by the pha-

sic electrodermal response to stress for a signal level

analysis.

Phasic electrodermal activity peaks have the fol-

lowing components: latency, rise time and half recov-

ery time, each which are expected to take 1-3 seconds

(Iveta et al., 2015). Therefore, a window length of 3

or 5 would likely only capture a segment of the full

peak. Event-based signal processing problems would

beneﬁt from a smaller window length as these anal-

yses are concerned with identifying speciﬁc points

in the signal. However, whole signal analysis, such

as identifying larger regions of the signal as associ-

ated with a period of physiological arousal, require

a larger window length. The window length must

capture enough of the signal to determine whether re-

gions have enough peaks to differ based on the state

of arousal the participant is in. This idea is exempli-

ﬁed in time-series heatmaps for window lengths 3 and

5 for both Biopac and E4 (Figs. 2 and 3) where the

heatmaps exhibit a “bar coding” visualization effect;

parallel windows are assigned to a different cluster,

rather than to the same cluster. This is particularly

apparent in the middle of the signal where, based on

the experimental protocol, the stress response would

be elicited and phasic peaks would occur. Therefore,

the clustering algorithm is likely clustering the differ-

ent components of the phasic peaks during this period,

clustering the latent and recovery periods with regions

of the signal that are associated with non-stressful or

baseline periods.

The identiﬁcation of atypical signals is less sensi-

tive to the size of the window length in noisy signals

that exhibit drift and dense phasic peaks throughout

the entire signal. At all window lengths participant 15

from dataset DS1Biopac was visibly different than the

rest of the participants used in the analysis (Fig. 2).

The discussion and ﬁndings of the window

length’s effect on the structure of feature data used in

this paper highlights the importance of understanding

the context of data used in machine learning experi-

ments and identifying an appropriate window length

for a signal-based or event-based analysis. Further, an

understanding of the context of the physiological pro-

cess of interest can yield a more interpretable analysis

and thus a greater understanding of the effects of input

parameters on the behaviour of algorithms used.

4.4 Heatmaps for Visualizing DBSCAN

Cluster Assignments in Time Series

The heatmap DBSCAN cluster assignment plots,

plotted as a time-series for each participant were an

effective measure to draw insights about how clus-

ters were being assigned across the signal, for each

participant, for each window length. In future stud-

ies, the visualization strategy of cluster assignments

can be enhanced by used ﬁxed colours between clus-

ter scatter plots and heatmaps in order to more easily

associate the scatter plots to the heatmap assignments.

4.5 Reﬂection of Datasets Used

The limitations of the dataset used are not an anomaly

in research; data scientists are often provided datasets

that are “messy” or with information missing about

how the data were collected or accessed. However,

this paper demonstrates that valuable insights can be

drawn from data of seemingly low quality. These in-

sights are not limited to the conclusions drawn thus

far; exploratory data analysis can bring to light the

importance of stringent research practices and provide

teams with issues in their practices that may otherwise

have not been apparent to them.

5 CONCLUSION

Unsupervised learning proved to be an effective ap-

proach to understanding the structure of EDA data

Unsupervised Electrodermal Data Analysis Comparison between Biopac and Empatica E4 Data Collection Platforms

351

when coupled with a time-series heatmap visualiza-

tion approach. The structure of the data was investi-

gated through clustering with k means and DBSCAN,

revealing that the structure of the data in both Biopac

and E4 were more suitable for DBSCAN, which is

robust to noise.

This demonstrates the ability of clustering to be

used to discover and characterize data structure even

when labels pertaining to activity descriptions are

missing, as was the case with this data.

Clustering with different sized window lengths

had a stark impact on the structure of 81 the data in

both Biopac and E4; at a higher window length (90

and 120 seconds), the data of two participants was

ﬂagged (the entire signal was clustered in a single

cluster, or was visually very different than the remain-

ing participants) and determined unusable.

ACKNOWLEDGEMENTS

We thank Dr. Kristal Thomassin and her students

for the use of the data from the Child Emotion

and Mental Health Lab at the University of Guelph

(https://www.childemotionlab.ca).

REFERENCES

Benedek, M. and Kaernbach, C. (2010). Decomposition of

skin conductance data by means of nonnegative de-

convolution. Psychophysiology.

Biopac (2019). EDA introductory guide. Technical Report

EDA Guide, Biopac Systems Inc., Goleta, California,

USA.

Bizzego, A., Battisti, A., Gabrieli, G., Esposito, G., and

Furlanello, C. (2019). pyphysio: A physiological sig-

nal processing library for data science approaches in

physiology. SoftwareX, 10:100287.

Boucsein, W. (2012). Electrodermal Activity. Springer Sci-

ence & Business Media.

Cecchi, S., Piersanti, A., Poli, A., and Spinsante, S. (2020).

Physical stimuli and emotions: EDA features analysis

from a wrist-worn measurement sensor. In 25th Int.

Workshop on Computer Aided Modeling and Design

(CAMAD). IEEE.

Charmandari, E., Tsigos, C., and Chrousos, G. (2005). Neu-

roendocrinology of the stress response. Annual Re-

view of Physiology, 67(1):259–284.

Chen, Y.-L., Kuan, W.-H., and Liu, C.-L. (2020). Compar-

ative study of the composition of sweat from eccrine

and apocrine sweat glands during exercise and in heat.

International Journal of Environmental Research and

Public Health, 17(10):3377.

Cho, D., Ham, J., Oh, J., Park, J., Kim, S., Lee, N.-K., and

Lee, B. (2017). Detection of stress levels from biosig-

nals measured in virtual reality environments using

a kernel-based extreme learning machine. Sensors,

17(10):2435.

Du, Q., Faber, V., and Gunzburger, M. (1999). Centroidal

voronoi tessellations: Applications and algorithms.

SIAM Review, 41(4):637–676.

Empatica (2020). E4 wristband: Real-time physiological

signals: Wearable PPG, EDA, temperature, motion

sensors. https://www.empatica.com/research/e4/.

Ester, M., Kriegel, H.-P., Sander, J., Xu, X., et al. (1996).

A density-based algorithm for discovering clusters in

large spatial databases with noise. In KDD-96 Pro-

ceedings, pages 226–231.

Greco, A., Valenza, G., Lanata, A., Scilingo, E., and Citi, L.

(2016). cvxEDA: a convex optimization approach to

electrodermal activity processing. IEEE Transactions

on Biomedical Engineering, pages 1–1.

Iveta, B., Igor, O., Michal, M., and Ingrid, T. (2015). Au-

tonomic nervous system in children with autism spec-

trum disorder. Cognitive Remediation Journal, 4.

Kirschbaum, C., Pirke, K.-M., and Hellhammer, D. H.

(1993). The ‘Trier Social Stress Test’ – a tool for in-

vestigating psychobiological stress responses in a lab-

oratory setting. Neuropsychobiology, 28(1-2):76–81.

Lowry, R. (1977). Active circuits for direct linear measure-

ment of skin resistance and conductance. Psychophys-

iology, 14(3):329–331.

MacQueen, J. (1965). Some methods for classiﬁcation and

analysis of multivariate observations. In Proceed-

ings of the Fifth Berkeley Symposium on Mathematical

Statistics and Probability, volume 1, pages 281–297.

University of California Press.

Makowski, D., Pham, T., Zen, Brammer, J. C., Le, D., Hung

Pham, Lesspinasse, F., Chuan-Peng Hu, and Sch

olzel,

C. (2020). Neurokit2: A python toolbox for neuro-

physiological signal processing.

Millington, G. W. M. and Graham-Brown, R. A. C. (2010).

Skin and skin disease throughout life. In Rook's Text-

book of Dermatology, pages 1–29. Wiley-Blackwell.

Posada-Quintero, H. F. and Chon, K. H. (2020). Innovations

in electrodermal activity data collection and signal

processing: A systematic review. Sensors, 20(2):479.

Sander, J., Ester, M., Kriegel, H.-P., and Xu, X. (1998).

Density-based clustering in spatial databases: The al-

gorithm gdbscan and its applications. Data Mining

and Knowledge Discovery, 2:169–194.

Schmelz, M., Schmidt, R., Bickel, A., Torebj

ork, H. E., and

Handwerker, H. O. (1998). Innervation territories of

single sympathetic c ﬁbers in human skin. Journal of

Neurophysiology, 79(4):1653–1660.

Tufte, E. R. (1991). Envisioning Information. Graphics

Press, Cheshire, Conneticut.

Watson, D., Clark, L. A., and Tellegen, A. (1988). Develop-

ment and validation of brief measures of positive and

negative affect: the PANAS scales. Journal of person-

ality and social psychology, 54(6):1063.

DATA 2022 - 11th International Conference on Data Science, Technology and Applications

352