Point Process Temporal Structure Characterizes Electrodermal Activity

Abstract

Electrodermal activity (EDA) is a read-out of the body's sympathetic nervous system measured as sweatinduced changes in the electrical conductance properties of the skin. There is growing interest in using EDA to track physiological conditions such as stress levels, sleep quality and emotional states. Standardized EDA data analysis methods are readily available. However, none considers two established physiological features of EDA: 1) sympathetically mediated pulsatile changes in skin sweat measured as EDA resemble an integrate-and-fire process; 2) inter-pulse interval times vary depending upon the local physiological state of the skin. Based on the anatomy and physiology that underlie feature 1, we postulate that inverse Gaussian probability models would accurately describe EDA inter-pulse intervals. Given feature 2, we postulate that under fluctuating local physiological states, the inter-pulse intervals would follow mixtures of inverse Gaussian models, that can be represented as lognormal models if the conditions favor longer intervals (heavy tails) or by gamma models if the conditions favor shorter intervals (light tails). To assess the validity of these probability models we recorded and analyzed EDA measurements in 11 healthy volunteers during 1 to 2 hours of quiet wakefulness. We assess the tail behavior of the probability models by computing their settling rates. All data series were accurately described by one or more of the models: two by inverse Gaussian models; five by lognormal models and three by gamma models. These probability models suggest a highly succinct point process framework for real-time tracking of sympatheticallymediated changes in physiological state.

A Model-Based Approach for Pulse Selection from Electrodermal Activity

“…We have previously shown that the inter-pulse interval distribution in EDA data follows an inverse Gaussian distribution [2] [3], which agrees with a model of the rise of sweat through the gland to the skin surface as an integrate-andfire process [4] [5]. This process is similar to the mechanism that underlies other point process events such as neuronal and cardiac action potentials, earthquakes, geysers, and volcanoes [6] [7].…”

“…This process is similar to the mechanism that underlies other point process events such as neuronal and cardiac action potentials, earthquakes, geysers, and volcanoes [6] [7]. We showed that the temporal structure in EDA favors right-skewed heavy tailed distributions, including the inverse Gaussian and even heavier tailed distributions such as the lognormal, due to the presence of sparse regions of EDA with low activity and long inter-pulse intervals [2]. This result makes possible the use of low-order models in EDA analyses and increases the signal-to-noise ratio.…”

“…Nevertheless, we observed a high degree of intersubject variability in the number of pulsatile events (pulses) despite all subjects being at rest for the same duration of time [2]. This is due to the inherent variability in background filling levels and spontaneous activity of sweat glands between people, as well as differing levels of noise inherent to different recording equipment and experimental conditions.…”

“…Preprocessing consisted of two major steps, 1) detecting and removing artifacts and 2) isolating the phasic component. Both have been described in previously in [2]. Because of the level of high frequency noise seen in the recording equipment used for the propofol data, those data were additionally low-pass filtered with a cutoff of 3 Hz after artifact removal.…”

“…To this end, the relationship between the properties of several right-skewed distributions helps identify the presence of noise in any subject cohort. For example, while the heavier-tailed inverse Gaussian and lognormal distributions are indicators of desired statistical structure, the flexibility of the gamma distribution in capturing the light tails likely to occur with random noise in a subject cohort allows it to function as a noise indicator [2]. By examining the patterns of goodness-of-fit of four distributions (inverse Gaussian, lognormal, gamma, exponential) across subjects in a cohort, we gauged the level of noise in the subject cohort and the set of pulses to extract.…”