Decomposition of skin conductance data by means of nonnegative deconvolution

Abstract

Skin conductance (SC) data are usually characterized by a sequence of overlapping phasic skin conductance responses (SCRs) overlying a tonic component. The variability of SCR shapes hereby complicates the proper decomposition of SC data. A method is proposed for full decomposition of SC data into tonic and phasic components. A two-compartment diffusion model was found to adequately describe a standard SCR shape based on the process of sweat diffusion. Nonnegative deconvolution is used to decompose SC data into discrete compact responses and at the same time assess deviations from the standard SCR shape, which could be ascribed to the additional process of pore opening. Based on the result of single non-overlapped SCRs, response parameters can be estimated precisely as shown in a paradigm with varying inter-stimulus intervals.

Figure 1

Diagram of two different sequences underlying an SCR (adapted from Edelberg, 1993, with friendly permission). If sweat ducts are filled to their limits, intraductal pressure will cause a hydraulic driven diffusion of sweat to the corneum resulting in a flat SCR (A). If intraductal pressure exceeds tissue pressure, the distal part of the duct and the pore will eventually open, which results in a peaked SCR (B).

Figure 2

Standard deconvolution and nonnegative deconvolution applied to model data of two differently shaped SCRs: (A) flat SCR, resulting from sweat diffusion only; (B) peaked SCR, resulting from sweat diffusion and additional pore opening. Standard deconvolution is depicted for two different impulse response functions (Bateman function with small and higher τ₂), while nonnegative deconvolution is only depicted for the latter impulse response function.

Figure 3

Sequence of the decomposition of SC data by means of nonnegative deconvolution. Given is a 60-s segment of raw SC data (A). Tonic SC activity is estimated based on inter-impulse data detected in the standard deconvolution of the raw SC data (B). Nonnegative deconvolution is applied to the phasic SC data (original SC data minus tonic SC activity) and single impulses and corresponding pore opening components are identified by means of segmentation of driver and remainder signal (C). The original SC data can finally be recomposed by superposition of its tonic and phasic components (D).

Figure 4

Normalized amplitudes of diffusion and pore opening components of SCRs, for all SCRs that occurred during the experiment. The solid and dashed lines show the median and the 25% and 75% percentiles for the size of the pore opening component for a given size of the diffusion component (in the range from 0 to 2 at bins of 0.25).

Figure 5

Event-related EDA in response to a startle probe for varying ISIs (4, 8, 16, and 32 s) depicted for 6 s relative to stimulus onset. Panel A shows the driver impulse for the 50 strongest responses. Panel B shows the average driver (impulse) and remainder data (*The remainder was amplified by factor 6 in order to be discriminable in a common scaling with the driver data). Panel C shows the SCR recomposed by diffusion and pore opening part.