During vigilance to painful stimuli: slower response rate is related to high trait anxiety, whereas faster response rate is related to high state anxiety

Abstract

A pathological increase in vigilance, or hypervigilance, may be related to pain intensity in some clinical pain syndromes and may result from attention bias to salient stimuli mediated by anxiety. During a continuous performance task where subjects discriminated painful target stimuli from painful nontargets, we measured detected targets (hits), nondetected targets (misses), nondetected nontargets (correct rejections), and detected nontargets (false alarms). Using signal detection theory, we calculated response bias, the tendency to endorse a stimulus as a target, and discriminability, the ability to discriminate a target from nontarget. Owing to the relatively slow rate of stimulus presentation, our primary hypothesis was that sustained performance would result in a more conservative response bias reflecting a lower response rate over time on task. We found a more conservative response bias with time on task and no change in discriminability. We predicted that greater state and trait anxiety would lead to a more liberal response bias. A multivariable model provided partial support for our prediction; high trait anxiety related to a more conservative response bias (lower response rate), whereas high state anxiety related to a more liberal bias. This inverse relationship of state and trait anxiety is consistent with reports of effects of state and trait anxiety on reaction times to threatening stimuli. In sum, we report that sustained attention to painful stimuli was associated with a decrease in the tendency of the subject to respond to any stimulus over time on task, whereas the ability to discriminate target from nontarget remains unchanged.

NEW & NOTEWORTHY During a series of painful stimuli requiring subjects to respond to targets, we separated response willingness from ability to discriminate targets from nontargets. Response willingness declined during the task, with no change in subjects’ ability to discriminate, consistent with previous vigilance studies. High trait anxious subjects were less willing to respond and showed slower reaction times to hits than low anxious subjects. This study reveals an important role of trait anxiety in pain vigilance.

INTRODUCTION

Attention bias as measured by reaction times to pain-related stimuli is associated with increases in pain in both acute pain and chronic pain syndromes (1–4). This increase in pain has been related to increased anxiety, and psychological activation or tense arousal (2, 3, 5, 6). Despite the importance of this clinical effect, vigilance to painful stimuli per se is rarely studied (7, 8). In clinical pain syndromes, vigilance is generally evaluated by self-report measures or attentional bias measures of response to stimuli that represent pain but are not painful (9–12).

Signal detection theory (SDT) allows vigilance tasks to be classified based on whether the decrease in performance during time on task (vigilance decrement) results from a decrease in discriminability or a more conservative response bias. SDT results in identification of response bias (B^″), the tendency to endorse a stimulus as a target, and discriminability (A′), the ability to discriminate target from nontarget. The vigilance decrement may result from a decrease in the subject’s tendency to respond to target stimuli or from diminished discrimination of target from nontarget (13–16). Reduced discriminability of visual and auditory stimuli over time on task generally occurs in vigilance tasks that require responses of ≥24 stimuli/min (15). During our continuous performance task (CPT), subjects discriminate between mild and moderate painful stimuli at ∼13 stimuli/min. Therefore, our primary hypothesis is that the vigilance decrement will result from a conservative shift in response bias, with no change in discriminability over time on task.

SDT-based analyses of painful stimuli have been controversial, particularly when measures of detection are equated with pain magnitude (17, 18). Because pain is subjective, it has been suggested that subjects may not use a consistent response bias during any session, much less across sessions, to report painfulness of stimulus-evoked sensations (19–21). Variation in criteria during task periods may confound the calculation of SDT metrics (19). SDT analysis of pain discrimination may be less susceptible to response bias variability (22, 23). Here we adopt a task of pain discrimination rather than pain detection that may allow separation of response bias and discriminability, consistent with studies of vigilance for nonpainful modalities, such as visual, auditory, and innocuous tactile stimuli (15, 16).

SDT metrics for detection of auditory and innocuous cutaneous stimuli have found that less-salient stimuli result in more conservative response biases and lower detectability of the stimulus (24–26). Therefore, we predict that a more conservative response bias will be related to lower stimulus salience.

Animals frequently reduce activity during exposure to a threatening environment (27–32). Reductions in movement, interpreted as increases in vigilance in response to a threating stimulus, such as during a confrontation with a gunman, have been reported to be exacerbated in high trait anxious individuals and in those with clinical anxiety disorders (33–37). In contrast, individuals with high state anxiety demonstrate more liberal response bias to detect pain in response to noxious stimuli (38–41). Which relationship will dominate in our CPT is unknown: an anxiety-associated liberal shift in response bias or a conservative shift in response bias often found during exposure to aversive or threatening stimuli (33, 34, 42–46).

METHODS

Subject Recruitment

We recruited 31 healthy volunteers from neighbors, staff, students, and trainees of the Johns Hopkins Hospital and University. Subjects who were enrolled included 17 women (age, 18–53 yr) and 14 men (age, 19–59 yr). The population was composed of 20 Caucasians [ 10 females), 5 African Americans (2 females), and 6 Asians (5 females). Of the 31 subjects enrolled, one subject (female, age 23 yr, Caucasian) withdrew from the study. To ensure subjects performed the CPT task attentively, we excluded an additional three subjects (1 male, age 31 yr, Caucasian; 1 female, age 23 yr, Asian; and 1 female, age 42 yr, Caucasian), with hit rates <0.57. Exclusion criteria included active neurological conditions, including the presence of chronic pain, psychiatric conditions, and medical conditions. All sessions took place between 8 AM and 6 PM Eastern Standard Time. All subjects gave written informed consent before the study, and all study procedures were approved and renewed yearly by the Johns Hopkins School of Medicine Institutional Review Board. Results of these experiments have been reported using traditional vigilance performance metrics (7). Here, we extend that analysis using signal detection theory and a novel multivariable exploratory analysis framework.

Stimuli Determination and Psychophysical Training

Painful heat stimuli were delivered with the Contact Heat-Evoked Potentials (CHEPs) thermode from Pathway (Medoc Advanced Medical Systems, Ramat Yishay, Israel). The CHEPs thermode has a contact surface area of 572.6 mm³ (circular surface with a diameter of 27 mm). To select the temperatures for CPT testing, pain ratings were obtained using a range of heat stimuli (42–51°C, in 1.0°C increments) that were applied to the left ventral forearm, with a plateau of 1.5 s and a rise rate of 70°C/s and a fall rate of 40°C/s. Each stimulus temperature was delivered two times for a total of 20 stimuli. During the training session, subjects rated the pain intensity of each of the 20 stimuli using a verbal scale anchored at 0 (no pain) and 10 (most intense pain imaginable). Mean pain intensity ratings closest to 2/10 (weakly painful stimuli) and 4/10 (moderately painful stimuli) were used to determine stimulus temperatures used for each subject during the CPT testing session. The target stimulus was the more painful of the two stimuli.

To guard against nociceptor fatigue and habituation, the probe was moved to an adjacent position for each stimulus. Specifically, stimulation began on the most proximal part of the lateral ventral forearm and was moved distally three times for the next three stimuli. Then the probe was moved to the proximal part of the medial ventral forearm and moved distally three times for the next three stimuli. This pattern was repeated for the entire train of stimuli. A total of eight positions on the ventral forearm were stimulated. The time between consecutive stimulations of the same area of the ventral forearm was at least 40 s.

To verify that subjects could determine the difference between the low and moderate pain temperatures, a brief training and verification session was performed. First, five stimuli from each of the two temperatures (low and moderate pain) were randomly delivered, and subjects were informed as to whether the applied stimulus was the lower or higher temperature. Next, 20 stimuli from each of the two temperatures were randomly delivered, and subjects were instructed to indicate whether the applied stimulus was the lower or higher temperature. An accuracy of ≥19/20 was required before proceeding to the CPT task. If a subject was unable to achieve this accuracy, temperatures were adjusted, and the subject was retrained and tested until they accurately identified ≥19/20 stimuli per temperature.

CPT Testing

A modified CPT-pain paradigm was used to investigate the underlying mechanisms of sustained attention to pain as illustrated in Fig. 1 The paradigm included four 10-min blocks of 120 stimuli consisting of 84 weakly painful nontarget stimuli and 36 moderately painful target stimuli. Stimuli were applied with 35°C serving as baseline, a rise rate of 70°C/s and a fall rate of 40°C/s, 1.3-s plateau durations, and randomly variable interstimulus intervals between 4.25 and 4.75 s. In addition, stimuli were randomized within each block to prevent order effects. Subjects were instructed to press a mouse button (Medoc Response Unit, Ramat Yishai, Israel) each time they perceived the moderate pain stimulus (target) and to withhold response to the low pain stimulus (nontarget).

Figure 1.

A: session overview, the order of the continuous performance task (CPT) and habituation sections was counterbalanced across subjects. B: subjects complete four blocks of the painful CPT. Each block consisted of 120 stimuli, 36 that were target (high intensity) stimuli distributed randomly. All stimuli were 1.3 s in duration with a random interstimulus interval (ISI) between 4.25 and 4.75 s. The dotted line in the lower panel is the threshold temperature for pain as established for subjects individually (see methods) (7).

Habituation Blocks

Subjects completed a habituation task to verify that outcomes were driven by attention, as opposed to mechanisms of habituation in the absence of a task. The order of presentation of the habituation task versus the CPT was randomized and counterbalanced across subjects. In all cases, the CPT-pain and habituation tasks were separated by a 40-min window, in which subjects completed self-report questionnaires (see psychological data). This arrangement separates the two protocols, and both are delivered after the same training stimuli. The habituation task included two, 2.5-min blocks consisting of 26 stimuli (13 weakly and 13 moderately painful stimuli) with identical parameters to the CPT-pain task. Stimuli were applied with 35°C serving as baseline, a rise rate of 70°C/s and a fall rate of 40°C/s, 1.3-s plateau durations, and randomly variable interstimulus intervals between 4.25 and 4.75 s. During the habituation blocks, subjects were not required to differentiate between the two stimulus types and did not have access to the mouse button.

Ratings

Following each CPT and habituation block, ratings for pain intensity, pain unpleasantness, stimulus-related salience, tense arousal, and attention to pain versus something else were obtained (47–49). For many of these variables, the average value and the change over time on task were calculated [(Block 3 + Block 4) − (Block 1 + Block 2)]/(Block 1 + Block 2). Pain intensity and pain unpleasantness were assessed using a separate manual visual analog scale (VAS), which was anchored at 0 (no pain) and 10 (most intense pain imaginable) for pain intensity and at 0 (no pain) and 10 (most unpleasant pain imaginable) for pain unpleasantness (50). The relationship of pain intensity and unpleasantness was explained to the subjects by relating the following passage: “There are two aspects of pain which we are interested in measuring: the intensity; how strong the pain feels and the unpleasantness, or how disturbing the pain is for you. The distinction between these two aspects of pain might be made clearer if you think of listening to a sound, such as a radio. As the volume of the sound increases, I can ask you how loud it sounds, or how unpleasant it is to you. The intensity of pain is like loudness; the unpleasantness of pain depends not only on intensity, but also on other factors which may affect you. These are scales for measuring each of these two aspects of pain. Although some pain sensations may be equally intense and unpleasant, we would like you to judge these two aspects of your pain (or the temperatures you feel) independently.” Specific instructions were given to the subjects regarding how to rate salience. Specifically, the form stated: “Salience is described as the ability for a stimulus to capture attention. On a scale of 0–10, with 0 being the ABSENCE of salience and 10 being the THE MOST SALIENT STIMULUS IMAGINABLE, how salient was the painful stimulus?” Tense arousal was measured using adjectives from the tense arousal subscale of the Activation-Deactivation Checklist (51). Subjects were instructed to “Please use the rating scale next to each word to describe your feelings at this moment.” Five adjectives (jittery, intense, clutched-up, tense, and fearful) were rated using a four-point scale anchored at 1 (definitely do not feel) and 4 (definitely feel). Tense arousal was averaged over three ratings recorded 1) before the start of the task, 2) after subjects completed training on the two stimulus temperatures, and 3) after verification of correct identification of the two stimulus temperatures.

Following the final CPT and habituation blocks, additional measures were captured regarding the subject’s attention to internal thoughts and concerns [mind-wandering (MW)], which was rated on a seven-point scale, where 1 indicates never and 7 indicates always (48). The explicit definition of mind-wandering used is thoughts unrelated to a participant’s present sensory environment (48). In this case the definition is with regard to performing CPT with painful thermal stimuli.

Psychological Data

To identify psychological factors that may contribute to pain and performance outcomes during sustained attention to pain, we had subjects fill out various psychological scales and questionnaires. These included: STAI—State and Trait Anxiety Inventory (52), and the Activation-Deactivation Checklist (51). All questionnaires were completed during a 40-min block that occurred between CPT and habituation testing (Fig. 1A).

Signal Detection Theory Analysis

When considering metrics of vigilance decrement, measures such as detection rate include both the detectability of the signal and the bias of the observer in responding to stimulus trials (13, 53). Signal detection theory (SDT) allows one to separate the signal detection process into a measure of the observer’s discriminability of a stimulus and the observer’s response criterion, also called the response bias (β) (54). In SDT, perceptual sensitivity, the observer’s ability to discriminate between signal and noise (or, in our case, target and nontarget stimuli, d′). Although several measures from SDT are available to calculate response bias or criterion, prior studies assessing the performance of various criterion metrics in vigilance have suggested using the criterion metric C (16). However, in our case, given the probability that hits (H) and false alarms (F) are not normally distributed, we used complementary measures based on the area of the calculated receiver operator curve (55). Specifically, A′, the corresponding measure for d′, is calculated by $A\mathrm{\prime }=\frac{1}{2}+\frac{(H -\mathrm{ }F)(1\mathrm{ }+\mathrm{ }H -\mathrm{ }F)}{4H(1\mathrm{ }-\mathrm{ }F)}$, when H ≥ F, and B^″, the corresponding measure for β, is calculated by B^″$\mathrm{ }=\mathrm{ }\frac{H\left(1\mathrm{ }-\mathrm{ }H\right)-F(1\mathrm{ }-\mathrm{ }F)}{H\left(1\mathrm{ }-\mathrm{ }H\right)+F(1\mathrm{ }-\mathrm{ }F)}$, when H ≥ F. SDT has a long history of contributions to vigilance research, allowing us to compare the performance of our observers to prior research on tasks in other sensory modalities (13, 53).

Statistical Analysis

All statistical analyses were conducted using R version 3.6.3. Data from all variables were assessed for normality using the Shapiro–Wilk test. Departures from normality were not substantial, and the sample size of 27 subjects (three subjects were removed from the data analysis because of poor task performance) justified the use of nontransformed data (56). Variables showing significant departures from normality were evaluated for the presence of outliers, defined as values more than three median absolute deviations from the median (57). We defined poorly performing subjects as those with average performance below chance level for 144 total target trials (cumulative probability ≥ 0.95; 82 hits out of 144 targets or 57% hit rate).

Effects of order of habituation and CPT on suprathreshold pain intensity ratings throughout the various phases of the protocol were communicated previously (7). In summary, reduction of pain intensity and unpleasantness ratings occurred only during habituation trials, not during the CPT trials.

We tested for a vigilance decrement over time as well as the effects of intrinsic attention to pain and mind-wandering on task performance using a linear mixed model (LMM), with the factor of time represented by the four blocks in the protocol (58). We included in the LMMs the intrinsic attention to pain (IAP) measure across the CPT (IAP scores: only something else = 1, mostly something else = 2, mostly pain = 3, only pain = 4), trait anxiety, state anxiety, and a Likert scale for mind-wandering (MW) (48). In the LMM, each outcome variable (e.g., number of hits or false alarms, hit or false alarm response time, response bias, or discriminability) was modeled as a within-subject factor of time, between-subject fixed-effects factors of trait and state anxiety, IAP and MW, and subject as a random effect. The effect of time was explicitly modeled by comparing block 1 with the other three blocks. This has the effect of reducing the number of possible pairwise comparisons and is warranted by extensive prior literature examining the vigilance decrement in other sensory modalities (59). This model was compared to a simple repeated-measures analysis of variance (RM-ANOVA) model assuming compound symmetry, a model including an autoregressive correlation structure for time, and a model for each test. For A′, the ordered model was superior by Akaike's Information Criteria, Bayesian Information Criteria, and log-likelihood tests (P < 0.0026). For B^″, there was no clear superior model. Considering the experimental design, the ordered model is the most appropriate for statistical hypothesis testing with regard to the effect of time.

Model selection in linear mixed model (LMM) or general linear model analysis is a nondeterministic process; multiple models with similar fixed factor structures are statistically indistinguishable when the number of variables in the model surpasses three. Although statistical and scientific principles can be applied to the model selection problem, no rigorous exploratory method currently exists to determine the most likely model given the experimental data. Guiding principles in model selection include parsimony, various measures of information criteria, significant improvement in model fit, and supporting evidence from prior scientific studies (58, 60, 61). Balancing type 1 error rates and statistical power in complex LMMs is contentious, with some authors arguing for testing maximal models and other authors arguing for parsimony and demonstrating that maximal models lack statistical power to support their reported effects (62, 63). Although this discussion occurred within the framework of confirmatory models and random-effects structures, parsimony is an equally valid principle for exploratory analysis when considering the effects of multiple interacting fixed effects on a particular outcome. To bridge this gap in model selection methods, we implemented a permutation algorithm for model selection in the LMMs in this paper.

Specifically, we evaluated the effects of the set of interacting fixed-effects time, trait anxiety, state anxiety, IAP, and MW on each outcome (number of hits or false alarms, hit or false alarm response time, response bias, or discriminability). We posited a minimal fixed-effects model structure of time plus at least one other input variable and populated the model space for each outcome variable up to and including the maximal model. In the case of psychological variables, the maximal fixed-effects part of the model is Time ×TraitAnxiety × StateAnxiety × MW × IAP. In the case of perceptual characteristic variables, the maximal model is Time ×StimulusSalience × PainIntensity × PainUnpleasantness. Fixing the model structure by requiring “time” to be present in all the models is justified scientifically, as our primary interest is in sustained attention or vigilance (64). Holding constant parts of the model structure with a priori hypotheses effectively reduces the search space of the model (fewer multiple comparisons or colloquially, forking paths; 65, 66). After generating all non-redundant models that fulfilled our model criteria, we tested all models and summarized the significant factors over the model space we had populated.

We include in our results the summary tables reporting the proportion of model terms significant at P < 0.05 and trending at P < 0.10 (Supplemental Tables S1–S12). Furthermore, we summarize the results of this exploratory multivariable analysis by describing the model that is a combination of all terms, where that term is significant at P < 0.05 in a majority of the model space. When no models have a particular term as significant at P < 0.05, we discard that term. In the rare case, where a fixed-effect factor term is compelling enough to investigate in the absence of statistical significance (e.g., evaluating the effect of state anxiety in a model that includes trait anxiety), we justify discussing that result based on prior evidence from previously published peer-reviewed studies. We report F-stat values for the fixed effects derived from LMMs using the anova function in R-base and P values derived using corrected degrees of freedom from the Kenward–Roger correction, whereas R² values are derived using the Kenward–Roger approach (67–69).

To evaluate the strength and direction of significant fixed-effect terms in the parsimonious final models, we used Pearson’s partial correlation analysis. Specifically, we constructed multivariable partial correlation models including the particular outcome variable. All bivariate and partial correlation models were tested using the ppcor package in R (70). This package solves the multivariable correlation problem by inverting the covariance matrix, solving all partial correlations simultaneously. This avoids the necessity of specifying a hierarchical model. Because we evaluated many variables at each task block, over time on task, we corrected for the repeated-measures correlation with a t test using the effective degrees of freedom. For figures, we used the R package ggscatter to create bivariate scatterplots with corresponding 95% confidence curves for relationships of interest (Figs. 3–5).

Figure 3.

Pearson correlations between hits (A) and response bias (B) with pain intensity, and hit reaction time and stimulus salience (C) across all blocks of the continuous performance task (CPT). The gray area indicates the 95% confidence interval.

In summary, this novel multivariable approach involves five steps: 1) Enumeration of all possible models within a constrained model structure (e.g., here as a time series). 2) Evaluation of the overall proportion of factors and interactions (i.e., terms) in the maximal model that are significant, where the maximal model refers to the fixed-effects part of the model that in all hypothesized variables and their interactions are included in the model. 3) Using a stepwise model construction approach, where the order of term introduction depends on only the strength of the relationship inferred through level and proportion of significance. 4) Testing the relative increase in variance explained by each term introduction step of the model construction. The final model is determined by testing the significance of the introduction of each term. 5) Using the construction of partial correlation models starting with the maximal model and removing nonsignificant variables to determine the direction and magnitude (e.g., approximate effect size) of each relationship.

RESULTS

Vigilance and Responsivity Decrement

Vigilance decrement is indicated by the decrease in hits over time, in a CPT with a painful target. During the CPT, hits significantly decreased during time on task across task blocks [F_3,78 = 5.79; P = 0.0013; comparisons of block 1 with blocks 2, 3, and 4 were significant (P ≤ 0.035); Fig. 2A]. The time factor explained significant variance in hits in the final linear mixed model (Time R² = 0.15; 95%CI = 0.052–0.33). False alarms also significantly decreased with time on task (F_3,78 = 3.56; P = 0.018; comparison of block 1 with block 3 was significant (P ≤ 0.0092); Fig. 2B). The time factor explained significant variance in false alarms in the final linear mixed model (Time R² = 0.12; 95%CI = 0.037–0.29). Applying signal detection theory (SDT), discriminability (A′) between the two painful stimuli in the CPT did not change during time on task (F_3,69 = 1.11; P = 0.35; Fig. 2C). In contrast, response bias (B^″) became significantly more conservative (increased) during time on task (F_3,78 = 8.69; P = 4.9 × 10⁻⁵); comparisons of block 1 with blocks 2, 3, and 4 were highly significant (P ≤ 0.00041; Fig. 2D). The time factor explained significant variance in response bias in the final linear mixed model (Time R² = 0.25; 95%CI = 0.12–0.43). During time on task in a CPT with a painful target, subjects, on average, adopted a more conservative response criterion, resulting in a vigilance decrement in both hits and false alarms (fewer responses over time on task). A spaghetti plot following each of the participants’ measurements across blocks for these same data is presented in Supplemental Fig. S1 (see https://doi.org/10.6084/m9.figshare.13227425.v3). These results echo the increase in misses over time on task we previously reported for these data (7).

Figure 2.

Vigilance and Signal Detection Theory metrics over the four blocks of continuous performance task (CPT) with a painful target. A: hits over time on task. B: false alarms over time on task. C: discriminability (A′) over time on task. D: response bias (B^″) overtime on task. Error bars represent SEM. *P < 0.05, **P < 0.01, and ***P < 0.001.

At no time during the CPT were discriminability and response bias correlated (R < |0.29|; P ≥ 0.14), and the change in response bias and discriminability during the CPT also were not correlated with each other (R = 0.070; P = 0.73), supporting the independence of these measures.

Response Bias during the CPT Was Influenced by Pain Intensity and Stimulus Salience

For the multivariable linear mixed-model analysis, we separately modeled the effects of perceptual characteristics and psychological factors on vigilance or signal detection theory measures.

In the LMM for hits, the main effects of time (F = 4.5, P = 0.0062), stimulus salience (F = 8.0, P = 0.0057), pain intensity (F = 4.6, P = 0.037), and the interaction of pain intensity and pain unpleasantness (F = 9.0, P = 0.0043) were all significant (Table 1; Supplemental Table S1; all Supplemental material is available at https://doi.org/10.6084/m9.figshare.12824558). The final model explained significant variance in hits R² = 0.34 (95%CI = 0.25–0.53). In a four-variable partial correlation model including hits, pain intensity, pain unpleasantness, and stimulus salience, only pain intensity was positively correlated with hits (R = 0.27; P = 0.015). In a reduced three-variable partial correlation model including hits, pain intensity, and stimulus salience, hits were positively correlated with pain intensity (R = 0.36, P = 0.00087), but negatively correlated with stimulus salience (R = −0.22, P = 0.049). Pain unpleasantness was only positively correlated with hits in a three-variable model including hits, pain unpleasantness (R = 0.25, P = 0.025), and stimulus salience (R = −0.18, P = 0.12). Finally, the bivariate correlation between hits and pain intensity remained significant (R = 0.31, P = 0.0051; Fig. 3A). In summary, in all models accounting for pain unpleasantness or stimulus salience, hits were strongly related to pain intensity. The relationship between pain intensity and hits mediated a negative relationship of hits with stimulus salience, whereas accounting for stimulus salience moderated the positive relationship of hits with pain unpleasantness.

Table 1.

Linear mixed-model stepwise selection process for main and interaction effects of stimulus salience, pain intensity, and pain unpleasantness

Outcome	Model Comparison	F-Stat	Adjusted Degrees of Freedom	P Value	Model R2
Hits	1. Hits∼1vs. Hits∼1+Time	5.19	3, 78	0.0025*	0.17 (0.063–0.34)
	2. Hits∼1+Timevs. Hits∼Time × Salience	2.02	4, 86.2	0.098+	0.23 (0.14–0.42)
	3. Hits∼Time × Saliencevs. Hits∼Time × Salience+PainInt × PainUnpl	5.71	3, 43	0.0022*	0.34 (0.25–0.53)
Hit reaction time (HRT)	1. HRT∼1vs. HRT∼1+Time+PainInt	2.73	4, 82.1	0.034*	0.12 (0.043–0.29)
	2. HRT∼1+Time+PainIntvs. HRT∼1+Time+PainInt+Salience	3.53	1, 88.7	0.063+	0.15 (0.069–0.33)
False alarm reaction time (FRT)	1. FRT∼1vs. FRT∼1+Time	2.67	3, 78	0.054+	0.093 (0.024–0.26)
	2. FRT∼1+Time vs. FRT∼1+Time+PainInt × PainUnpl	1.95	3, 79.8	0.13	0.13 (0.064–0.31)
Response bias (B″)	1. B″∼1vs. B″∼1+Time	12.2	3, 78	1.3 × 10−6*	0.32 (0.18–0.49)
	2. B″∼1+Timevs. B″∼1+Time+PainInt × PainUnpl	3.27	3, 62.1	0.027*	0.35 (0.23–0.51)
	3. B″∼1+Time+ PainInt × PainUnplvs. B″∼1+Time+PainInt  × PainUnpl+Salience	4.94	1, 74.7	0.029*	0.39 (0.27–0.55)

The final model is in bold in the model comparison column. Degrees of freedom are approximated using the Kenward–Rogers approach. In the P value column, +one-tailed, whereas *two-tailed significance. Model R² is reported as the mean and 95% confidence interval. B^″, response bias; FRT, false alarm reaction time; HRT, hit reaction time; PainInt, pain intensity; PainUnpl, pain unpleasantness.

There was no significant relationship between pain intensity, pain unpleasantness, or stimulus salience on either false alarms or pain discriminability (A′) (Supplemental Tables S2 and S3).

In the LMM for response bias (B^″), the main effects of time (F = 9.9, P = 1.3 × 10⁻⁵), stimulus salience (F = 5.2, P = 0.025), pain intensity (F = 4.1, P = 0.047), and the interaction of pain intensity and pain unpleasantness (F = 7.7, P = 0.0074) were all significant (Table 1; Supplemental Table S4). The final model explained significant variance in B^″ R² = 0.39 (95%CI = 0.27–0.55). In a four-variable partial correlation model including B^″, pain intensity, pain unpleasantness, and stimulus salience, stimulus salience was positively correlated with B^″ (R = 0.26; P = 0.015), whereas pain intensity trended to a negative correlation with B^″ (R = −0.20, P = 0.072). In a reduced three-variable partial correlation model including B^″, pain intensity, and stimulus salience, response bias was positively correlated with stimulus salience (R = 0.25, P = 0.024), but negatively correlated with pain intensity (R = −0.33, P = 0.0027). Pain unpleasantness was only negatively correlated with response bias in a three-variable model including B^″, stimulus salience (R = 0.22, P = 0.039), and pain unpleasantness (R = −0.27, P = 0.014). Finally, the bivariate negative correlation between B^″ and pain intensity remained significant (R = −0.26, P = 0.021; Fig. 3B). In summary, in all models accounting for pain unpleasantness or stimulus salience, response bias was negatively correlated with pain intensity. The relationship between pain intensity and response bias mediated a positive relationship with stimulus salience, whereas accounting for stimulus salience moderated a negative relationship of hits with pain unpleasantness. This is similar to, though opposite to, the pattern observed with hits, which follows, as B^″ is derived from both hit and false alarm rates.

In the LMM for hit reaction time (HRT), the main effect of pain intensity was significant (F = 12.0, P = 0.00081), whereas stimulus salience trended toward significance (F = 3.6, P = 0.061) (Table 1; Supplemental Table S5). The final model explained significant variance in HRT (R² = 0.15; 95%CI = 0.069–0.33). In a three-variable partial correlation model including HRT, pain intensity, and stimulus salience, stimulus salience was positively correlated with HRT (R = 0.35; P = 0.0017), whereas pain intensity was negatively correlated with HRT (R = −0.30, P = 0.0064). Although the bivariate positive correlation between HRT and stimulus salience remained significant (R = 0.25, P = 0.023; Fig. 3C), HRT and pain intensity were no longer significantly correlated in the bivariate relationship (R = −0.18, P = 0.10). Therefore, stimulus salience mediates the negative correlation between HRT and pain intensity. In blocks where heat pulses were reported as having greater pain intensity, subjects demonstrated faster reaction times to correctly identified targets. In contrast, in blocks where heat pulses were reported as more salient stimuli, subjects demonstrated slower reaction times, when controlling for pain intensity.

Finally, in the LMM for false alarm reaction time (FRT), the main effect of time was significant (F = 3.5, P = 0.019), whereas neither pain intensity (F = 2.8, P = 0.098) nor pain unpleasantness (F = 0.11, P = 0.74) was significant, but the pain intensity by pain unpleasantness interaction trended toward significance (F = 3.79, P = 0.055; Table 1; Supplemental Table S6). The final model explained significant variance in FRT (R² = 0.13; 95%CI = 0.064–0.31). In a three-variable partial correlation model including FRT, pain intensity, and pain unpleasantness, pain intensity was negatively correlated with FRT (R = −0.26; P = 0.020), but pain unpleasantness was not significantly correlated with FRT (R = 0.17, P = 0.13). The bivariate negative correlation between FRT and pain intensity only trended toward significance (R = −0.20, P = 0.067). Therefore, in blocks where painful heat was reported as having greater pain intensity, subjects demonstrated faster reaction times to incorrectly identified nontargets. Controlling for pain unpleasantness strengthened the relationship between FRT and pain intensity.

State and Trait Anxiety Have Opposing Effects on Response Bias

The second set of LMMs, we assessed the relationships of psychological measures such as state and trait anxiety, intrinsic attention to pain and mind-wandering on vigilance or signal detection theory measures. The mean state anxiety score for our experimental group was 26.7 ± 6.6 (SD), whereas the mean trait anxiety score was 32.5 ± 7.7 (SD). The means and standard deviations do not suggest any significant pathological anxiety disorders in our healthy experimental group compared with normative data from healthy adults where the mean state anxiety is 35.6 ± 10.5 (SD) and trait anxiety 34.9 ± 9.2 (SD) (52).

In the LMM for hits, the main effects of time (F = 5.19, P = 0.0025), state anxiety (F = 5.49, P = 0.028), and trait anxiety (F = 14.3, P = 0.00091) were significant (Table 2; Supplemental Table S7). The final model explained significant variance in hits (R² = 0.32; 95%CI = 0.19–0.51). In a three-variable partial correlation model including hits, state, and trait anxiety, trait anxiety (R = −0.46; P = 0.000019) was negatively correlated, whereas state anxiety (R = 0.31, P = 061) was positively correlated with hits. The bivariate negative correlation between hits and trait anxiety remained significant (R = −0.37, P = 0.00062; Fig. 4A), whereas the bivariate correlation between hits and state anxiety was no longer significant (R = 0.092, P = 41). Therefore, trait anxiety mediates the relationship between hits and state anxiety.

Table 2.

Linear mixed-model stepwise selection process for main and interaction effects of trait anxiety, state anxiety, mind-wandering, and intrinsic attention to pain

Outcome	Model Comparison	F-Stat	Adjusted Degrees of Freedom	P Value	Model R2
Hits	1. Hits∼1 vs. Hits∼1+Time	5.19	3, 78	0.0025*	0.17 (0.063–0.34)
	2. Hits∼1+Time vs. Hits∼1+Time+TraitAnx	7.98	1, 25	0.0091*	0.24 (0.12–0.42)
	3. Hits∼1+Time+TraitAnxvs. Hits∼1+Time+TraitAnx+StateAnx	5.49	1, 24	0.028*	0.32 (0.19–0.51)
False alarms (FAs)	1. FAs∼1 vs. FAs∼1+Time	3.56	3, 78	0.018*	0.023 (0.005–0.14)
	2. FAs∼1+Time vs. FAs∼1+Time+TraitAnx	3.20	1, 25	0.086+	0.15 (0.062–0.34)
	3a. FAs∼1+Time+TraitAnxvs. FAs∼1+Time+TraitAnx × MW × IAP	1.57	6, 67.7	0.17	0.24 (0.17–0.46)
	3b. FAs∼1+Time+TraitAnxvs. FAs∼1+Time × MW × IAP+TraitAnx	1.06	12, 73.4	0.41	0.26 (0.22–0.50)
Discriminability (A′)	1. A′∼1 vs. A′∼1+Time × TraitAnx × StateAnx × MW	1.14	31, 56.7	0.33	0.38 (0.40–0.67)‡
Response bias (B″)	1. B″ ∼1 vs. B″ ∼1+Time	12.2	3, 78	1.26x10-6*	0.32 (0.18–0.49)
	2. B″ ∼1+Timevs. B″ ∼1+Time+TraitAnx	4.71	1, 25	0.040*	0.35 (0.22–0.52)
	3. B″ ∼1+Time+TraitAnx vs. B″ ∼1+Time+StateAnx × IAP+TraitAnx	2.82	3, 61.8	0.046*	0.40 (0.28–0.57)
	4a. B″ ∼1+Time+StateAnx × IAP+TraitAnxvs. B″ ∼1+Time × StateAnx × MW + StateAnx × IAP+TraitAnx	1.15	11, 73.5	0.34	0.47 (0.39–0.65)
	4b. B″ ∼1+Time+StateAnx × IAP+TraitAnxvs. B″ ∼1+Time+StateAnx × IAP + MW × IAP+TraitAnx	3.31	2, 43	0.046*	0.47 (0.35–0.64)
	5a. B″ ∼1+Time × StateAnx × MW + StateAnx × IAP+TraitAnxvs. B″∼1+Time × StateAnx × MW+StateAnx × IAP+MW × IAP+TraitAnx	2.99	1, 87.8	0.087+	0.49 (0.42–0.67)
	5b.B″∼1+Time+StateAnx × IAP+MW × IAP +TraitAnxvs. B″∼1+Time × StateAnx × MW+StateAnx × IAP+MW × IAP+TraitAnx	0.96	10, 71.2	0.49	0.49 (0.42–0.67)

The final model is in bold in the model comparison column. Degrees of freedom are approximated using the Kenward–Rogers approach. In the P value column, +one-tailed, whereas *two-tailed significance. Model R² is reported as the mean and 95% confidence interval. A′, discriminability; B^″, response bias; FAs, false alarms; FRT, false alarm reaction time; HRT, hit reaction time; IAP, intrinsic attention to pain; MW, mind-wandering; StateAnx, state anxiety; TraitAnx, trait anxiety. ^‡95% confidence interval is not accurate, a result of overfitting and model instability.

Figure 4.

Pearson correlations of trait anxiety and hits (A), response bias (B), and hit reaction times (C and D) over all blocks of the continuous performance task (CPT). The gray area indicates the 95% confidence interval.

In the LMM for false alarms, the main effect of time (F = 3.56, P = 0.018) was significant, whereas trait anxiety was not significant (F = 3.20, P = 0.086; Table 2; Supplemental Table S8). The final model explained significant variance in false alarms (R² = 0.15; 95%CI = 0.062–0.34). The bivariate negative correlation between false alarms and trait anxiety was significant (R = −0.31, P = 0.0055).

In the LMM for discriminability (A′), no main effects or interactions among mind-wandering, intrinsic attention to pain (IAP), state anxiety or trait anxiety were significant (Table 2; Supplemental Table S9). The only significant term from the exploratory analysis that fit our criteria was the four-way interaction among the effects of time, mind-wandering, state anxiety, and trait anxiety. However, the model comparison test between the model containing this term and the baseline (intercept-only) model was not significant (F = 1.14, P = 0.33). This result, in addition to the nonsignificant model intercept (t = 0.49, P = 0.63), is strong evidence that this significant four-way interaction is the result of overfitting the model in the context of a relatively small sample size.

In the LMM for response bias (B^″), the main effects of time (F = 8.69, P = 4.9 × 10⁻⁵), state anxiety (F = 4.49, P = 0.044), and trait anxiety (F = 10.2, P = 0.0041) were all significant, whereas the interactions between state anxiety and IAP (F = 2.78, P = 0.099), and mind-wandering and IAP (F = 2.93, P = 0.090) were not significant (Table 2; Supplemental Table S10). The final model explained significant variance in B^″ (R² = 0.47; 95%CI = 0.35–0.64). In a five-variable partial correlation model including B^″, IAP, mind-wandering, state anxiety, and trait anxiety, trait anxiety (R = 0.42, P = 0.00012) and mind-wandering (R = 0.25, P = 0.028) were positively correlated, whereas state anxiety was negatively correlated with B^″ (R = −0.32, P = 0.0038). There was no relationship between IAP and B^″ (R = −0.0095, P = 0.93). Furthermore, after removing IAP from the model, trait anxiety (R = 0.43, P = 0.000080) and mind-wandering (R = 0.25, P = 0.024) were positively correlated, whereas state anxiety was negatively correlated with B^″ (R = −0.32, P = 0.0036). In a reduced three-variable partial correlation model, trait anxiety (R = 0.40, P = 0.00025) was positively correlated, whereas state anxiety was negatively correlated with B^″ (R = −0.28, P = 0.012). Finally, the bivariate correlation between B^″ and trait anxiety remained significant (R = 0.31, P = 0.0046; Fig. 4B), but not the bivariate correlations between B^″ and state anxiety (R = −0.10, P = 0.36) or mind-wandering (R = 0.17, P = 0.13). Therefore, trait anxiety mediates the relationship among B^″, state anxiety, and mind-wandering. The strongest relationships were between B^″ and trait and state anxiety, such that trait anxiety mediates the relationship between B^″ and state anxiety.

In the LMM for hit reaction time (HRT), the main effects of IAP (F = 4.45, P = 0.038) and trait anxiety (F = 10.01, P = 0.0042) were significant, as were the interactions between mind-wandering and IAP (F = 3.99, P = 0.049) as well as IAP and trait anxiety (F = 6.43, P = 0.013; Table 3; Supplemental Table S11). The final model explained significant variance in HRT (R² = 0.28; 95%CI = 0.18–0.47). In a four-variable partial correlation model, HRT was positively correlated with IAP (R = 0.39, P = 0.00036) and trait anxiety (R = 0.50, P = 0.000002), but not with mind-wandering (R = −0.11, P = 0.32). In a reduced three-variable partial correlation model, HRT was positively correlated with IAP (R = 0.40, P = 0.00020) and trait anxiety (R = 0.50, P < 0.000001). Bivariate correlations between HRT and IAP (R = 0.46, P = 0.000014) and trait anxiety (R = 0.54, P < 0.000001) remained significant (Fig. 4C and 5A). The partial correlation models showed the interaction between mind-wandering and IAP was not relevant for the relationships among HRT, IAP, and trait anxiety. Therefore, the positive relationships of HRT and IAP, as well as those of HRT and trait anxiety, are largely independent and robust.

Table 3.

Linear mixed model stepwise selection process for main and interaction effects of trait anxiety, state anxiety, mind-wandering, and intrinsic attention to pain

Outcome	Model Comparison	F-Stat	Adjusted Degrees of Freedom	P-Value	Model R2
Hit reaction time (HRT)	1. HRT∼1vs. HRT∼1+Time+TraitAnx	3.42	4, 75.4	0.013*	0.15 (0.062–0.34)
	2a. HRT∼1+Time+TraitAnxvs. HRT∼1+Time+TraitAnx × IAP	5.53	2, 79.4	0.0057*	0.24 (0.14–0.43)
	2b. HRT∼1+Time+TraitAnxvs. HRT∼1+Time+MW × IAP+TraitAnx	2.99	3, 64.2	0.037*	0.23 (0.14–0.44)
	3a. HRT∼1+Time+TraitAnx × IAPvs. HRT∼1+Time+TraitAnx × IAP+MW × IAP	2.28	2, 47.6	0.11	0.28 (0.18–0.47)
	3b. HRT∼1+Time+MW × IAP+TraitAnxvs. HRT∼1+Time+TraitAnx × IAP+MW × IAP	6.40	1, 77.8	0.013*	0.28 (0.18–0.47)
False alarm reaction time (FRT)	1. FRT∼1vs. FRT∼1+Time	2.67	3, 78	0.054+	0.093 (0.024–0.26)
	2. FRT∼1+Timevs. FRT∼1+Time+StateAnx	4.98	1, 25	0.035*	0.15 (0.057–0.33)
	3. FRT∼1+Time+StateAnxvs. FRT∼1+Time+StateAnx × TraitAnx	3.07	2, 23	0.066+	0.26 (0.15–0.50)
	4a. FRT∼1+Time+StateAnx × TraitAnxvs. FRT∼1+Time+StateAnx × TraitAnx+TraitAnx × IAP	2.47	2, 90.1	0.091+	0.28 (0.18–0.49)
	4b. FRT∼1+Time+StateAnx × TraitAnxvs. FRT∼1+Time+StateAnx × TraitAnx × MW	1.77	4, 19	0.18	0.48 (0.35–0.73)
	5a. FRT∼1+Time+StateAnx × TraitAnx +TraitAnx × IAPvs. FRT∼1+Time+StateAnx × TraitAnx × MW+TraitAnx × IAP	1.50	4, 19	0.24	0.46 (0.35–0.70)
	5b. FRT∼1+Time+StateAnx × TraitAnx × MWvs. FRT∼1+Time+StateAnx × TraitAnx × MW+TraitAnx × IAP	1.91	2, 88	0.15	0.46 (0.35–0.70)

The final model is in bold in the model comparison column. Degrees of freedom are approximated using the Kenward–Rogers approach. In the P value column, +one-tailed, whereas *two-tailed significance. Model R² is reported as the mean and 95% confidence interval in parentheses. A′, discriminability; B^″, response bias; FAs, false alarms; FRT, false alarm reaction time; HRT, hit reaction time; IAP, intrinsic attention to pain; MW, mind-wandering; StateAnx, state anxiety; TraitAnx, trait anxiety.

Figure 5.

Pearson correlations between intrinsic attention to pain and choice reaction times over all blocks of the continuous performance task (CPT) (A and B). Pearson correlation between trait anxiety and intrinsic attention to pain (C). The gray area indicates the 95% confidence interval. RT, reaction time.

Finally, in the LMM for false alarm reaction time (FRT), the interaction between trait anxiety and IAP was significant (F = 4.12, P = 0.045), whereas the main effects of state anxiety (R = 4.18, P = 0.053) and time (F = 2.18, P = 0.097) did not reach significance (Table 3; Supplemental Table S12). The interaction between state and trait anxiety was not significant (F = 3.24, P = 0.085). The final model explained significant variance in FRT (R² = 0.28; 95%CI = 0.18–0.49). In a four-variable partial correlation model, FRT was positively correlated with IAP (R = 0.22, P = 0.050), state anxiety (R = 0.25, P = 0.027), and trait anxiety (R = 0.205, P = 0.070). Bivariate correlations between FRT and IAP (R = 0.28, P = 0.013), state anxiety (R = 0.36, P = 0.0010), and trait anxiety (R = 0.37, P = 0.00064) were significant (Figs. 4D and 5B). Trait anxiety was significantly positively related to IAP (R = 0.24, P = 0.028; Fig. 5C). The moderating effect of IAP and state anxiety on the relationship between FRT and trait anxiety is revealed by the reduction in the correlation coefficient when state anxiety and IAP are added to the correlation model of FRT and trait anxiety. This interrelationship is clear in the positive correlation between trait anxiety and IAP.

DISCUSSION

Using signal detection theory (SDT), we separated response bias (B^″) from discriminability (A′) during a continuous performance task (CPT) with painful stimuli (25, 53, 71). We found a conservative shift in response bias over time and no change over time in discriminability. These results are consistent with our hypothesis, which was based upon previous findings from vigilance toward stimuli across nonpainful sensory modalities with similar event rates (13, 14). We found an unexpected relationship of response bias to salience of the stimulus such that more salient stimuli were associated with a more conservative response bias. In prior reports of vigilance to visual, auditory, and nonpainful cutaneous stimuli, salience was related to a more liberal response bias; i.e., the more salient the stimulus, the greater the hit and false alarm rate (13, 25, 26, 53, 72, 73). Finally, we found a divergence in the relationships of state and trait anxiety with response bias. Specifically, subjects with high trait anxiety scores had a more conservative response bias, whereas subjects with high state anxiety demonstrated a more liberal response bias, which may be related to the differential effects of state versus trait anxiety on attentional bias to threatening stimuli (34, 42, 44, 74–76).

Conservative Response Bias Shift with Time on Task, No Change in Discriminability

We hypothesized that the relatively slow rate of stimuli in our protocol would result in a more conservative response bias over time on task (13, 14, 77). We found evidence that the vigilance decrement was associated with a conservative shift in bias. This decrement occurred during the first 10-min block and then remained stable during the following three 10-min blocks (Fig. 2). In contrast, there was no change in discriminability over time. Therefore, the change in response bias but not discriminability resulted in a progressive reduction of both false alarms and hits during time on task.

The relatively slow conduction velocity of Aδ- and C-fibers decreases the temporal resolution of signal processing in the nociceptive system requiring a relatively slow rate of stimulus presentation compared with the rate of auditory or visual stimuli traditionally used in vigilance studies (13, 78). Furthermore, painful stimuli that are applied in rapid succession will show temporal summation if given at a rate >0.33 Hz (79, 80). Therefore, all painful stimuli that would be perceptually independent and distinct (where one stimulus does not affect the next) would likely result in a similar behavioral response profile, where the response bias would become more conservative over time on task (13).

Stimulus Salience and Pain Intensity Have Opposing Effects on Response Bias during the CPT

Studies of auditory, visual, and nonpain tactile stimuli report that salience of the target is associated with a more liberal response bias (25, 26, 73, 81). Accordingly, we expected a similar relationship for painful stimuli. This is the first study to apply SDT to a classical vigilance task for discrimination between target and nontarget painful stimuli. We found that stimulus salience was positively correlated with response bias, meaning subjects who reported the stimuli as more salient had a lower overall response rate in contrast to stimuli in other sensory modalities. In contrast to stimulus salience, greater reported pain intensity correlated with a more liberal response bias. Therefore, although pain intensity and stimulus salience are positively related, within the context of the partial correlation model, greater reported stimulus salience was associated with a more conservative response bias, whereas greater pain intensity was associated with a more liberal response bias. The increased response rate to stimuli reported as more intensely painful is consistent with prior findings in vigilance studies in the auditory and visual domain. Similarly, a more conservative response bias to more salient negatively valenced stimuli is consistent with salient threatening stimuli promoting avoidance and retarding attention disengagement when stimuli are longer than 1 s in duration (82, 83). In this sense, measuring both self-reported pain intensity and stimulus salience allowed us to parse the separate contributions of subjective salience and intensity of the painful heat stimuli.

Inverse Relationships of State and Trait Anxiety on Response Bias

If the processing of acute painful stimuli is similar to that of other aversive stimuli, then trait anxiety should be related to greater avoidance of aversive stimuli, better performance on tasks that require withholding of a response, and failure to disengage from aversive stimuli (44, 45, 76, 84–87). The effect of greater trait anxiety is to exacerbate hypervigilance to aversive stimuli, frequently to the detriment of task performance (27, 28, 30, 37, 88). Therefore, during our CPT involving discrimination of painful stimuli, subjects with higher trait anxiety could be expected to show greater avoidance of the task, fewer false alarms, and dwell for a prolonged time on each aversive stimulus. This effect may be seen in our data as a more conservative response bias in subjects with higher trait anxiety (Fig. 5B). Further supporting this contention, subjects with high trait anxiety had longer RTs during correctly identified targets compared to subjects with low trait anxiety.

In fact, the RTs of subjects with high trait anxiety were not different between false alarms and hits, whereas subjects with low trait anxiety had significantly faster RTs during hits compared with false alarms. In contrast, high state anxiety is associated with subjects’ enhanced attentional capture of threatening stimuli. This suggests higher state anxiety during the session should be associated with a more liberal response bias (42, 44, 46, 74, 75, 83, 89, 90). We saw no significant correlation of greater state anxiety with a more liberal response bias unless both trait and state anxiety were included in a model with response bias. In that model, a clear association between greater state anxiety and a more liberal response bias was revealed.

It is unclear what behavioral mechanism supports the effect of trait anxiety on response bias in our paradigm. For example, we did not find any relationship between our discriminability measure and trait anxiety, as has been reported for detection of fearful faces (91). A likely scenario is that the response bias shift during our protocol is driven by a maladaptive hypervigilant state in high trait anxious subjects resulting from the processing demands of discriminating the painful stimuli (27, 30, 88). Future research could explore whether the relationship between a conservative response bias and a hypervigilant state might be related to “attentive immobility” as reflected in changes in alterations in muscle tone and heart rate in subjects performing a CPT with a painful target. The relationship between trait anxiety and conservative response bias shift might result from failure of attentional disengagement from the aversive stimulus consistent with findings of previous studies (44, 45, 85, 87). Furthermore, the relationship between state anxiety and a more liberal response bias should track with enhanced attentional capture as has been found in studies of healthy populations (75, 89, 92, 93). In contrast to previous studies of painful stimuli, we found no effect of anxiety on discriminability in contrast to previous studies using SDT to analyze effects of anxiety on response bias and discriminability (35, 39–41, 94–96).

Potential Neural Mechanisms

Recent studies in humans and nonhuman primates have found specific brain regions involved in neural correlates of discriminability, response bias, and report criterion of stimuli (97–100). Specifically, neural correlates of discriminability have been determined to be found in primary sensory areas. Whereas, evidence for response bias correlates have been localized to the putamen for painful stimuli and dorsolateral prefrontal cortex (DLPFC) for visual stimuli (98, 99). The localization of response bias shifts to the putamen and DLPFC may be under the influence of the amygdala, which is an important mediator of attentional processing, particularly as related to aversive stimuli. These results may implicate connections of the DLPFC, with the amygdala and putamen as mediating the individual variability in response bias to painful stimuli (101, 102).

Limitations

Our results show a strong relationship between anxiety and response bias, as well as solid evidence of a response bias shift during time on task in our CPT with painful stimuli. However, these results are subject to the same limitations as described in methods and interpretation of previous reports of SDT analysis of the detection of painful stimuli (17–19, 103, 104). Although most early studies applied SDT to the problem of pain detection, the inherent subjectivity of the label of a stimulus as painful has led many authors to question the validity of the method (105–107). Response bias instability has also been pointed out to contaminate sensitivity indices of pain (19). In the present study, we largely avoid these questions by applying SDT only to pain discrimination, not detection (22, 23).

The results of the exploratory investigative framework reported in this study, which represents the stepwise creation of multivariable testable models, requires testing and comparison of findings to an independent data set. Therefore, our resulting models of each vigilance performance metric must be considered preliminary pending testing of those models in an independent data set.

Conclusions

Using SDT to evaluate the vigilance decrement in a CPT with a painful target, we found the vigilance decrement was characterized not by a change in discriminability, but a conservative shift in response bias, similar to results from vigilance tasks presenting stimuli at a similar rate (13). High trait anxiety was associated with a more conservative response bias (fewer responses), whereas high state anxiety was associated with a more liberal response bias, but only when modeled with trait anxiety.

GRANTS

This work was supported by National Institute of Neurological Disorders and Stroke grant R01-NS107602 (to F.A.L.) and by JHNPRI (to F.A.L., N.M.E., and T.J.M.).

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

N.M.E., J.-H.C., A.K., J.D.G., and F.A.L. conceived and designed research; N.M.E. performed experiments; T.J.M., J.-H.C., M.I.S., and F.A.L. analyzed data; T.J.M., J.-H.C., O.J.B., A.K., J.D.G., and F.A.L. interpreted results of experiments; T.J.M. and M.I.S. prepared figures; T.J.M. and F.A.L. drafted manuscript; T.J.M., N.M.E., J.-H.C., M.I.S., O.J.B., A.K., J.D.G., and F.A.L. edited and revised manuscript; T.J.M., N.M.E., J.-H.C., M.I.S., O.J.B., A.K., J.D.G., and F.A.L. approved final version of manuscript.