Is it time to rethink motion artifacts?

0 downloads 0 Views 1002KB Size Report
expressions, gestures, whole body movements, etc.), and speech. Such data .... were not restricted, (b) the data was collected during a real-life situation with real ...
Is it time to rethink motion artifacts? Temporal Relationships between Electrodermal Activity and Body Movements in Real-life conditions Ryad Chellali, Shannon Hennig Pattern Analysis and Computer Vision Istituto Italiano di Tecnologia (IIT) Genova, Italy [email protected], [email protected] Abstract— This work investigates the temporal dynamics between electrodermal activity (EDA) and gestures during reallife public speaking. Wrist movements and EDA were recorded from speakers during a high-stakes public speaking event. The extreme values theory, more typically used for detecting relatively rare events in financial markets or flood calculations, was adapted and combined with continuous wavelet transform to automatically detect significant temporal events from the derivatives of the movement and EDA time series. The literature describing motion artifacts in EDA predicts that detected movement events should precede or co-occur with EDA. In contrast, we also observed a second case in which EDA changes preceded movement (approximately 40% of cases). These results suggest that humans may move or gesture in response to affective changes, which has clear implications for anyone studying arousal and gesture ‘in the wild.’ Keywords — Real world conditions, temporal dynamics, continuous wavelet transform, extreme values theory, electrodermal activity, movement artifacts

I.

INTRODUCTION

A key challenge in affective computing is understanding the relationships between the measurable signals produced during social interaction and cognitively demanding tasks. Numerous measurements can now be taken in a non-invasive manner during real-world activities. Such measurements include electrodermal activity (EDA), movements (i.e., facial expressions, gestures, whole body movements, etc.), and speech. Such data take the form of continuous time series upon which individual events occur at specific moments in time. With the advent of wearable, wireless EDA sensors (e.g., [1]), increasingly researchers record EDA as a way to monitor arousal levels and changes in affective experiences during real-world interactions and tasks. During such activities, a person’s movements and EDA are expected to fluctuate over time. The temporal dynamics of and events within these recorded signals are known to contain information about a person’s affective experience. Further, both arousal levels (and thus EDA readings) and movement can interact with the dynamics of a social situation. EDA measurements are increasingly being recorded in real-world situations that inherently involve human movement

(e.g., measuring arousal during driving [2], social interaction [3], and children interacting with technology [4]). This is in contrast to traditional methodologies that considered movement to be a source of artifacts in the EDA signal (i.e., motion artifacts). Such methods carefully minimize movement during EDA data acquisition. This presents a methodological tradeoff; non-moving participants produce EDA signals that are easier to interpret with existing methods, however many of the activities of interest to the affective sciences community require movement. Further, restricting movement and gesture can alter affective states and social interactions. The cognitive load literature is one source of evidence suggesting that we do not fully understand the temporal interactions between movement and arousal. Specifically EDA has been shown to increase with increased cognitive load [5] (as well as with increases in attention, emotional state, exposure to novel salient stimuli, and arousal level more generally), however gestures are also thought to influence how one copes with cognitive load [6]. For example, gesturing has been shown to assist with organizing one’s thoughts and facilitate verbal expression [6]–[9] whereas restricting gesture has been shown to negatively impact verbal expression and spatial memory performance [10], [11]. Further, gesturing while speaking appears to reduce the effects of high cognitive load as suggested by Shi et al’s work documenting cases in which average EDA, typically high in the presence of high cognitive load, was observed to be on average lower when participants were allowed to gesture [5]. This suggests that complex interactions may exist between gesture and EDA, with movement potentially influencing or reacting to underlying processes related to EDA. Relatively little is known about how EDA and movement signals change over time. This is understandable given that observation of motion artifacts in the EDA signals [12] has led to researchers to adopt protocols explicitly avoid motion during recording. Such motion artifacts are thought to take one of two forms, as illustrated in Fig. 1. In the first case, movement precedes an observed EDA event, presumably because some aspect of the movement also triggered a change in the sudomotor system. Such an artifact typically follows within 1-3 seconds of the movement [12]. In the second type of artifact, sharp changes in EDA and movement occur nearly simultaneously. This can occur when a sensor is too loose and movements cause the sensor to shift relative to the skin.

Alternatively, sharp movements cause changes in the pressure of the sensor against the skin impacting EDA readings. In contrast to these two cases, the third possibility is the case in which EDA events precede movement, which would not be considered a motion artifact. This possibility is not discussed in the literature to our knowledge, but is particularly interesting in light of Shi et al’s work [5] which suggests that gesture may play a role in how people respond to high arousal states. If this is the case, movements following changes in EDA during unconstrained tasks could potentially be related to attempts to regulate one’s arousal state through movement. Three possible temporal sequences of EDA and Movement

II.

METHODOLOGY

A. Data The data used for these analyses was collected during a real-world high stakes public-speaking task (thesis defenses). Data was successfully recorded from 7 speakers (females = 4) who wore a Q sensor (www.affectiva.com) on a single wrist. This sensor recorded both EDA and 3 degrees of acceleration at a sample rate of 32 Hz. The data consists of approximately 20 minutes of public speaking followed by 10-15 minutes of answering questions from the audience for each speaker (see [21] for additional detail). Consistent with the literature, the inter-subject variability of EDA was high. As illustrated in figure 2, speakers 3, 4 and 5 showed similar trends, namely EDA increased at the beginning of the talk, lowered at end of the talk, and rose again during the question and answer session immediately following the presentation. In contrast, speaker 1 showed a different profile with a constantly increasing EDA. Such variability prevents direct comparison of raw EDA levels and instead relative changes in the time series were considered.

Fig. 1. Relative temporal alignment of movement events versus EDA events with regard to whether they are considered to be movement artifacts or not.

Existing knowledge of the temporal dynamics of EDA and gesture are limited by unresolved methodological challenges related to how to appropriately analyze the complex EDA signal. Specifically, the EDA signal is composed of a slower moving, general trend (skin conductance level, SCL) upon which a series of discrete skin conductance responses (SCR) is superimposed [12]. Analyzing these two components is challenging, particularly given the idiosyncratic response patterns between people and the variable delays in the signals. A common way to cope with these challenges is to ignore the moment to moment temporal dynamics and instead compare average EDA levels or response rates over blocks of time, rather than as a time series. The tradeoff of such approaches is a loss of information about temporal dynamics between EDA and other temporal signals, such as movement. In contrast, in this work we aim to explore whether EDA events also precede movement events (column three in Fig 1) from data collected in real life conditions in which participants were allowed to freely move while speaking. This work contrasts with other work in that (a) gestures and movement were not restricted, (b) the data was collected during a real-life situation with real communication and cognitive demands, and (c) our time series analyses do not average data over time. During this analysis, a novel method of detecting EDA and movement events was developed using an adaptation of the extreme values theory and continuous wavelet transform.

EDA (uS)

Inter-subject variability of EDA

Fig. 2. The EDA (uS) of four of the seven participants during a high stakes public speech (approximately 20 minutes) and question/answer session (of variable length). The vertical bar illustrates the transition to the Q/A session.

B. Preparing the data From the raw EDA data, a low-pass filter (set at 10Hz) was used to filter out noise. The accelerometer data was combined by taking the x, y, and z components at any given point in time, i, such that Mov (i ) =

Acc

x

(i )

2

+ Acc

y

(i )

2

+ Acc

(i ) . (1) 2

z

3

For each of the resulting two time series (i.e., filtered EDA and the combined movement time series generated from (1)), the derivative was calculated over a sliding window of 100 samples (3.125 seconds given 32 Hz sample rate) and a step of 1 sample. The derivative was selected because the rate of change in these signals is thought to be more informative than the absolute values. These derivatives were then normalized to facilitate comparisons between the signals and between subjects, as illustrated in Fig. 3 and 4. C. Extreme Values Theory From these derivatives, a method for automatically detecting significant changes, or “events”, was needed. For this purpose, a modified use of the extreme values theory (EVT) was selected. EVT originally was developed within the

branch of statistics that investigates the behavior and occurrence of extreme values in time series. Typically it has used in the insurance and banking sectors for risk evaluation, however recently, the use of EVT was extended to hydrology for studying relatively infrequent events, such as flooding [18]. Filtered EDA signal versus derivative

Fig. 3. A representative 93.75 second (3000 samples) section of one participant’s filtered EDA signals (black) and normalized deriviative (red).

Raw accelerations and derived movement signal

that for a given distribution, the extreme values contained in that distributions’ tail will converge toward a specific distribution. In other words, the distribution of the set of values found in a tail is expected to belong to one of the distributions in the set of known GEV distributions (i.e., either the Gumble, Fréchet or a Weinbull distributions) allowing one to identify the normalized values that exceed a given threshold, given that as the threshold approaches the maximum values of the variable in question, the distribution of the values above the threshold is predictable as described by Pickands [14]. Frequently the distribution of such values is modeled using the Generalized Pareto Distribution (GPD) [15], as described by Pickands. In the following sections, we outline how the GPD can be used to extract noteworthy EDA and movements events. Specifically, an event Et is assumed to occur at time t if its value is greater than a threshold u, using the following method. Assuming that Xi (EDA or movement derivative in our case) is a set of independent and identically distributed (i.i.d.) variables, let M n = max { X 1 , ... X n } , if there exists a sequence of constants an > 0 and bn such that:

P

M n − bn ≤z ⎯ ⎯→ G(z),n ⎯ ⎯→ +∞ an ,

(2)

for z ∈ R in which G is a non-degenerate general extreme value distribution (e.g., Gumble, Fréchet or a Weinbull distributions). Fig. 4. A representative 93.75 second (3000 samples) section of one participant’s raw acceloremeter signals in red, green, and blue (x, y, z) and derived movement signal, calculated with (1), in black.

In our case, the EVT is not used in a canonical way: based on our working assumption that meaningful changes in EDA and movement are relatively rare in comparison to the underlying signal, though they do occur, at varying rates, throughout the recordings, we use the EVT as a clustering tool to isolate meaningful changes (abnormal behaviors) in both EDA and movements. Further we suspect that these detectable changes are related to underlying changes in phenomena of interest, such as a speaker’s cognitive load and/or affective state.

Then for i=1,…,n, we have:

P [X i ≤ z | X i > u] ⎯ ⎯→ H (z),u ⎯ ⎯→ uend ,

(3)

That is to say, the probability P of having an extreme event z, follows a binomial distribution H. For the special case of modeling the tails, namely when values exceed a given threshold, the limiting distribution is the generalized Pareto distribution:

H (y ) = 1− 1+ ξ

y −μ

σ

−1/ξ

,

(4)

Specifically, we adapted and applied the extreme values theory, namely we use the POT (peak over threshold) approach [19] to extract and identify automatically thresholds above which EDA and movement signals are considered to be extreme events. These thresholds are used to filter the results from the continuous wavelet transform. This allows us to extract the salient events for subsequent analysis regarding the temporal distance between the events and whether an EDA event precedes or follows a specific movement event.

with ( μ , σ , ξ ) the three parameters approximating the tail’s location, scale and shape respectively. Relevant to our purposes, this method was used to automatically determine an optimal threshold for all subjects above which EDA and movement were classified as significant event. Indeed, the optimal threshold u should be the one leading to a fixed H. All data above this cutoff threshold was considered to be an event, and those below it were considered to be a non-event, and excluded from further study at this time.

D. Adaptation of the Extreme Values Theory The extreme values theory identifies relatively unusual behaviors by identifying extreme deviations from the median of the probability distributions. One powerful tool for understanding these extremes, or tails of distributions, is the generalized extreme value distribution (GEV) [13]. Within this branch of statistics is the Fisher–Tippett theorem stating

E. Threshold selection In our specific case, the original time series are continuous and the extreme values independence hypothesis should not be valid. Indeed, we have no guaranty that the actual distribution of the peaks converges toward a GEV and any found threshold will not correspond to extreme events definition. To circumvent this difficulty, a de-clustering step is performed as

suggested in R-package POT [20]: extremes are taken from appropriately chosen contiguous blocks [see 19 for more details]. In our tests, the same threshold u was obtained for all subjects with identical values for ( μ , σ , ξ ) . It was determined by varying the value of the threshold u in (3) until both the scale and shape of the tail’s distribution became consistent with one of the asymptotic distribution, as predicted by the Fisher-Tippet theorem. Selecting the cutoff in this manner appears to be functionally valid given that little difference is seen in the number of detected events for cutoff values at or below the selected cutoff threshold, whereas above it, variance begins to increase (and so does the number of identified events). As illustrated in figure 5, the distribution’s scale and shape remain relatively constant until the threshold u exceeds 0.4 (highlighted in red) value identified using our method for this set of data.

Movement events detected using proposed inspired from the literature. Specifically, we tookmethod advantage of existing knowledge of the derivative and theoretical shape of skin conductance responses presented in [17] summarized in the following equation:

Fig. 6. Detected movement events (vertical pink bars) superimposed on the filtered movement signal (in purple) for subject 4.

EDA events detected using proposed method

Threshold selection using Fisher-Tippet theorem

Fig. 7. The detected EDA peaks (vertical pink bars) and the low’pas filtered EDA levels (in purple) for subject 4 before the final filtering described in (6) in which dected events are only validated if they match the known shape of skin conductance responses (i.e., stared example above).

EDA (t ) = e

Fig. 5. The threshold for subject 4: the shape and the scale remain more or less constant until u reaches 0.4, at which point variance increases.

F. Peaks detection in EDA and movements time series After applying corresponding thresholds to the derivative of EDA and movements signals, we have the segments of the original signals containing the extreme EDA and movements events. Following that, a wavelet-based peak detector was used to extract timely accurate peaks in the obtained time series (filtered EDA, derivative of movement, and derivative of EDA). Indeed, the POT does not give exact times the extreme events take place and the chosen wavelet-based technique allows detecting peaks avoiding time and scale artifacts. Classical peak detection techniques use amplitude variations within a fixed size window as the main criteria for determining whether a point is a peak. In contrast, Pan Du and colleagues [16] consider the “shape” as an additional criterion. Specifically, they developed a continuous wavelet transform (CWT) based technique that identifies peaks at different scales and amplitudes. Figures 6 and 7 illustrate the detected movement and EDA events, respectively, against with the filtered signals using our method with the data from the recorded doctoral student defenses. The last step of our algorithm is specific to EDA. We refine the EDA peak detection by using a theoretical model



t

τ1

−e



t

τ2

,

(5)

in which t refers to time. The expected SCR rise is described by τ1 =0.75, the decay by τ 2 = 2. Using this equation, we compared the detected EDA peaks to this theoretical EDA shape and excluded any detected peaks with a derivative that did not sufficiently correlated to the expected shape described in the literature (i.e., Fig 8a and (5)).

Fig. 8. On the left, the theoretical skin conductance response (EDA peak) shape described in the literature and equateion 5. On the right, the average detected peak shape detected with our methods from subject 4’s data.

G. Occurrence of EDA and Movements events After the final peak detection, two vectors remain: EDA and movement peaks. Our question regards whether EDA peaks ever precede movement peaks. As mentioned in the literature, when movements precede EDA peaks, the resulting EDA peaks are considered to be potential motion artifacts. In our data, we had been observing the reverse order: movements occurring after EDA peaks. In order to verify our observations of EDA preceding movement, we construct a distance matrix distMat(i,j) as follows:

, (6) where t (EDA i peak ) is the moment the ith EDA event occurs and t (MOV j peak ) is the moment the jth movement event

most movement peaks occurred 3.01 seconds after the EDA peak. Categorization of EDA-movement pairs

occurs. An example of the resulting matrix from one participant is given in Fig 9. For sake of visibility, the original distance matrix disMat (i, j ) is replaced by its exponential

e



disMat (i, j ) L

, with L = 50. Fig. 10. The 4 second windows before and after the EDA events. When movmenent occurs after EDA (right side of figure), the time difference is negative (EDA – movement); Motion artifacts follow the pattern on the left (movement before EDA).

TABLE I.

.

RATIO OF PAIR WITH EDA OCCURING FIRST BY SUBJECT

Number of movement events 4 seconds before and after EDA peaks

Fig. 9. Absolute distance matrix of the temporal differences between each EDA event (EDA event’s count along the X axis) and each individual movement event (1st, 2nd, … nth movement event along the Y axis). As the color gradiant become lighter (and approaches 1.0), the relative time between the 2 events decrease and approach being simutaneous. The black dotted line represent one specific combination, specifically the 100th detected movement event and the 85th detected EDA event were closely temporally aligned.

In Fig 9. EDA and movement events are observed to occur in close temporal succession to each other throughout the talk, however little information is provided by the matrix regarding the order of the events. To determine the relative differences between the temporal occurrence of individual EDA and movement events, we subtracted the time of each EDA event from any movement events that occurred within 4 seconds (before or after) the EDA event, as illustrated in Fig. 10. A histogram of the number of movement events that occurred before or after the associated EDA event (see Fig. 11a) was calculated. From this, a density function (Fig. 11b) was calculated to estimate the distribution of movement peaks relative to EDA peaks. The numeric counts are presented in Table 1. III.

RESULTS

From all 7 speakers, 797 movement peaks were detected with our method within the selected 4 seconds windows. Of these, 59.3% of EDA events occurred after a movement event (consistent with a movement artifact), however in 40.1% of cases, EDA events actually preceded movement. As illustrated in Fig. 11b, a bimodal distribution clearly is present in the density plot of detected events relative to the relative time difference between EDA and movement. For motion artifacts, movement peaks most frequently occurred 2.71 seconds before EDA, whereas when it followed EDA,

Subject ID

Movement first (# of pairs)

EDA first (# of pairs)

Percent being motion artifacts (movt first)

1

76

47

61.8 %

3

51

35

59.3 %

4

70

53

56.9 %

5

72

45

61.5 %

6

52

34

60.4 %

7

93

71

56.7 %

8

29

19

60.4 %

TOTAL

443

304

59.3 %

As predicted by the literature, motion artifacts were frequently observed in this real-life dataset (i.e., movement before EDA). The speaker with the largest percentage of motion artifacts was participant 1 with 61.8 % of her EDAmovement event pairs fitting this pattern; participant 7 had the lowest number of motion artifacts, 56.7%. The unique contribution of this work is that we also observed a smaller, but sizeable proportion of EDA events that occurred before a movement event. To the best of our knowledge, this is the first report of the relative percentage of EDA-to-motion pairs in real-world conditions. Specifically in approximately 40% of occasions in which EDA occurred within 4 seconds of a peak in wrist movement, the EDA event actually occurred before the movement. This cannot be explained as a movement artifact given that the EDA event had already occurred when the movement was observed. This percentage was relatively consistent across our 7 subjects, who otherwise showed very heterogeneous EDA profiles, with a range of non-motion artifacts being 38-44% of the total EDAmovement pairs. Further, when various capture windows were considered (i.e., 6 sec before and after, 1.5 seconds before and after EDA, etc.), similar clear bimodal distributions were observed.

Time differences between EDA-Movement event pairs

References

Fig. 11. Distributions of number of movement events occurring before and after EDA peaks sorted by the temporal difference between evetns.

IV.

DISCUSSION AND CONCLUSION

Given the trend towards more naturalistic data collection and the advent of wearable sensors, it is no longer realistic to presume that participant movement can be restricted during EDA recording. Here we proposed a novel method for detecting EDA and movement events and applied it to real-world data collected from seven people during a high-stakes public speaking task in which free movement was permitted. The detected EDA peaks were shown to be similar in shape to the description of skin conductance responses in the literature, suggesting that these are not outliers, but relevant information. Comparing the timing of the EDA and movement events revealed that in addition to the well-known movement artifacts, a second type of EDA-movement event was observed. Specifically, in 40% of EDA-movement pairs, EDA events preceded movement, which strongly suggests the existence of at least 2 phenomena and that the temporal dynamics between these cases differ (i.e., for the motion artifact case, movement occurs on average 2.17 before EDA, whereas in the new case presented here, movement occurs 3.01 seconds on average after EDA). In light of our findings from real-world data, the temporal order of changes in both EDA and movement will need to be carefully considered, because motion artifacts cannot explain nearly 40% of the data we observed. Future work is needed to better characterize and understand what is happening when EDA events closely precede wrist movements and to replicate this observation with additional data. While we are not saying that EDA predicts motion, only that it proceeds it in a significant number of cases, one hypothesis to be explored is whether gesture and movement is occurring in reaction to heightened arousal states perhaps as a regulatory mechanism or a means of expending energy. These results also indicate that the temporal order EDA and movement event should be carefully considered in future work utilizing EDA recording in the wild where movement is unconstrained.

[1] Ming-Zher Poh, N. C. Swenson, and R. W. Picard, “A Wearable Sensor for Unobtrusive, Long-Term Assessment of Electrodermal Activity,” Ieee Trans. Biomed. Eng., vol. 57, no. 5, pp. 1243–1252, May 2010. [2] J. A. Healey and R. W. Picard, “Detecting stress during real-world driving tasks using physiological sensors,” Intell. Transp. Syst. Ieee Trans., vol. 6, no. 2, pp. 156–166, 2005. [3] S. J. Guastello, D. Pincus, and P. R. Gunderson, “Electrodermal arousal between participants in a conversation: Nonlinear dynamics and linkage effects,” Nonlinear Dyn. Psychol. Life Sci., vol. 10, no. 3, pp. 365– 399, 2006. [4] I. Leite, R. Henriques, C. Martinho, and A. Paiva, “Sensors in the wild: exploring electrodermal activity in child-robot interaction,” in Proceedings of the 8th ACM/IEEE international conference on Human-robot interaction, Piscataway, NJ, USA, 2013, pp. 41–48. [5] Y. Shi, N. Ruiz, R. Taib, E. Choi, and F. Chen, “Galvanic skin response (GSR) as an index of cognitive load,” in CHI ’07 Extended Abstracts on Human Factors in Computing Systems, New York, NY, USA, 2007, pp. 2651–2656. [6] M. W. Alibali, S. Kita, and A. J. Young, “Gesture and the process of speech production: We think, therefore we gesture,” Lang. Cogn. Process., vol. 15, no. 6, pp. 593–613, 2000. [7] S. Goldin-Meadow, “The role of gesture in communication and thinking,” Trends Cogn. Sci., vol. 3, no. 11, pp. 419–429, Nov. 1999. [8] I. Cutica and M. Bucciarelli, “‘The More You Gesture, the Less I Gesture’: Co-Speech Gestures as a Measure of Mental Model Quality,” J. Nonverbal Behav., vol. 35, no. 3, pp. 173–187, Apr. 2011. [9] D. McNeill, “So you think gestures are nonverbal?.,” Psychol. Rev., vol. 92, no. 3, pp. 350–371, 1985. [10] K. J. Pine, H. Bird, and E. Kirk, “The effects of prohibiting gestures on children’s lexical retrieval ability,” Dev. Sci., vol. 10, no. 6, pp. 747–754, 2007. [11] E. Morsella and R. M. Krauss, “The role of gestures in spatial working memory and speech,” Am. J. Psychol., pp. 411–424, 2004. [12] M. E. Dawson, A. M. Schell, and D. L. Dilion, “The Electrodermal System,” in Handbook of psychophysiology, 3rd ed., J. T. Cacioppo, L. G. Tassinary, and G. G. Berntson, Eds. New York: Cambridge Univ Pr, 2007, pp. 159–181. [13] E. Omey, F. Mallor, and E. Nualart, “An introduction to statistical modelling of extreme values. Application to calculate extreme wind speeds,” Hogeschool-Universiteit Brussel, Faculteit Economie en Management, 2009. [14] J. Pickands III, “Statistical inference using extreme order statistics,” Ann. Stat., pp. 119–131, 1975. [15] C. Klüppelberg and A. May, “Bivariate extreme value distributions based on polynomial dependence functions,” Math. Methods Appl. Sci., vol. 29, no. 12, pp. 1467–1480, 2006. [16] P. Du, W. A. Kibbe, and S. M. Lin, “Improved peak detection in mass spectrum by incorporating continuous wavelet transform-based pattern matching,” bioinformatics, vol. 22, no. 17, pp. 2059–2065, 2006. [17] M. Benedek and C. Kaernbach, “Decomposition of skin conductance data by means of nonnegative deconvolution,” psychophysiology, Mar. 2010. [18] Fawcett, L. and Walshaw, D. “Improved Estimation for Temporally Clustered Extremes”. Environmetrics, 18, 2, pp. 173—188, 2007. [19] Chavez-Demoulin V. and Davison A. C. Modelling time series extremes. REVSTAT - Statistical Journal, 10(1), 109-133, 2012. [20] Mathieu Ribatet, “Generalized Pareto Distribution and Peaks Over Threshold”, R-Package http://r-forge.r-project.org/projects/pot/, 2012. [21] Hennig, S. (2011). Candidacy of physiological measurements for implicit control of emotional speech synthesis. Proceedings of the Doctoral Consortium. ACII 2011 Affective Computing and Intelligent Interaction. Memphis. Springer