Adaptive thresholding increases sensitivity to detect changes in the rate of skin conductance responses to psychologically arousing stimuli in both laboratory and ambulatory settings

Abstract

Psychophysiologists recording electrodermal activity (EDA) often derive measures of slow, tonic activity-skin conductance level (SCL)-and faster, more punctate changes-skin conductance responses (SCRs). A SCR is conventionally considered to have occurred when the local amplitude of the EDA signal exceeds a researcher-determined threshold (e.g., 0.05 μS), typically fixed across study participants and conditions. However, fixed SCR thresholds can preferentially exclude data from individuals with low SCL because their SCRs are smaller on average, thereby reducing statistical power for group-level analyses. Thus, we developed a fixed plus adaptive (FA) thresholding method that adjusts identification of SCRs based on an individual's SC at the onset of the SCR to increase statistical power and include data from more participants. We assess the utility of applying FA thresholding across two independent samples and explore age and race-related associations with EDA outcomes. Study 1 uses wired EDA measurements from 254 healthy adults responding to evocative images and sounds in a laboratory setting. Study 2 uses wireless EDA measurements from 20 children with autism in a clinical environment while they completed behavioral tasks. Compared to a 0.01, 0.03, and 0.05 μS fixed threshold, FA thresholding at 1.9% modestly increases statistical power to detect a difference in SCR rate between tasks with higher vs. lower subjective arousal and reduces exclusion of participants by up to 5% across both samples. This novel method expands the EDA analytical toolbox and may be useful in populations with highly variable basal SCL or when comparing groups with different basal SCL. Future research should test for reproducibility and generalizability in other tasks, samples, and contexts. IMPACT STATEMENTS: This article is important because it introduces a novel method to enhance sensitivity and statistical power in analyses of skin conductance responses from electrodermal data.

Keywords: Electrodermal activity (EDA); Research equity; Sensitivity; Skin conductance response (SCR); Statistical power.

Figure 1.

Example time course showing that higher SC is associated with larger short-term fluctuations in SC. Blue circles indicate rapid SC fluctuations greater than 0.05 μS, i.e., SCRs. The bottom left panel shows a portion of data with higher SCL during which there may be some spurious SCRs shown by the red arrows. The bottom right panel shows a portion of data with lower SCL in which deflections that could be potential SCRs were not detected.

Figure 2.

Each SCR can be characterized by several distinct and measurable features. Example values of a single SCR are in black. Red measurable features are based on EDA level and are in units of μS, and include (1) SC(onset) - the SC at a trough (low point) that precedes a peak (high point); the low and high points are identified computationally by calculating first and second temporal derivatives in widely used peak-detection algorithms; we also refer to SC(onset) as the prevailing SCL, (2) SC(peak) - the SC at the peak that follows a trough, (3) SCR amplitude - the difference in SC values between the peak and its preceding trough. Following the definition of Boucsein et al. (2012), we consider SCL to be the mean of SC over some time period, after ignoring portions of data comprising SCRs. The mean SC is the average SC over some time period that does include SCRs. The blue measurable features are in time units (e.g., seconds), and include (1) onset time - time of the SC(onset), (2) rise time - the time between SC(onset) and SC(peak), (3) peak time - the time of SC(peak), and (4) half recovery time - the time from SC(peak) to the time at which the SC has reduced to a level of (SC(peak) − 0.5×SCR amplitude). We propose a novel EDA metric: the skin conductance response amplitude percent (RAP), which is the SCR amplitude divided by the SC value at onset (i.e., the trough preceding the SCR peak). One of our main EDA outcomes is SCR rate, which is the number of SCRs per unit time. The Methods section provides more details on how these are calculated.

Figure 3.

In the exploratory portion of Study 1, we determined the optimal FA threshold level using curve-fitting to maximize effect sizes comparing the SCR rate during high vs. low arousal images and sounds in 254 healthy adults. Each figure shows two sets of data: unpleasant high arousal stimuli minus neutral low arousal stimuli (thick line) and pleasant high arousal stimuli minus neutral low arousal stimuli (thin lines). The vertical red lines in (a) and (b) show the optimal FA threshold level of 1.9% found by averaging each of the four optimal FA threshold levels. The lower plots show the percentage of the sample that are responders—who exhibited at least one SCR across all conditions at the given threshold—and the number of SCRs across all stimuli.

Figure 4.

FA thresholding (red solid line) yields slightly smaller error bars, higher goodness of fit (R²) in nearly all cases, and a more linear and monotonic relation between SCR rate and arousal ratings compared to traditional fixed thresholds of 0.01 μS (purple line at top), 0.03 μS (teal dotted line), and 0.05 μS (blue dashed line). Each of the 254 participants contributed five data points (X, Y pairs) to each subplot (i.e., 5 arousal ratings plus EDA data). The error bars are standard errors. The x-axis values were offset for display purposes only (not analysis) to avoid overlapping error bars.

Figure 5.

In Study 2, among 20 children with autism spectrum disorder, FA thresholding (a, red) yields higher effect sizes than traditional fixed thresholding (c, blue) in comparisons between high vs. low arousal tasks: a face-to-face conversation vs. puzzle. Results are consistent with and extend results from Study 1, suggesting consistency over the three sessions over eight weeks (thin, thicker, and thickest lines). The vertical red line in (a) shows the optimal threshold value 1.9% found in the exploratory portion of Study 1. Study 2 was treated as a test for replication and the effect size curves here did not influence or inform the optimization of FA threshold values found in Study 1. The lower plots in (b) and (d) show the percentage of responders who exhibited at least one SCR across all conditions at the given threshold and the number of SCRs across all stimuli. For the exploratory portion of Study 2, each effect size trace was curve-fitted to determine the FA threshold value that maximizes the effect size.