Better Safe Than Sorry: A Common Signature of General Vulnerability for Psychopathology

The nature of N/NE

Resulting from both genetic and environmental factors and their interactions, N/NE appears as a broad and stable disposition that leads one to appraise situations as threatening, to hold negative anticipations about oneself and the world, to experience negative mood states and emotions, and to show poor emotion regulation (Barlow, Ellard, et al., 2014; McCrae & Costa, 2003; Ormel et al., 2013; Ormel, Riese, & Rosmalen, 2012; Watson & Clark, 1984). N/NE is different from a negative affective state in response to environmental factors. In general, individuals with high trait N/NE show more affective instability; that is, they show more moment-to-moment affective fluctuations, even between feelings of different quality (Kuppens, Van Mechelen, Nezlek, Dossche, & Timmermans, 2007). Within a constructionist view on emotion (Barrett, 2017), N/NE should be seen as a valenced background state reflecting a rather long-term imbalance between the necessary and actual/anticipated resources to secure basic needs (e.g., growth, survival, reproduction) rather than as a categorical constructed emotion.

N/NE is associated with more vigilance toward potentially negative information; behavioral inhibition; a greater intolerance for uncertain, ambiguous, and uncontrollable situations; and a lower threshold for reacting with avoidance and escape responses to threat and negative information, including one’s own negative emotions (Barlow, Sauer-Zavala, et al., 2014; Gray & McNaughton, 1996). N/NE is also associated with an array of behavioral characteristics such as attentional (Shackman et al., 2016a, 2016b) and interpretational (Mathews, Ridgeway, Cook, & Yiend, 2007) bias; worrying, ruminating, and catastrophizing (Ehring & Watkins, 2008); enhanced symptom reporting (Van den Bergh, Witthöft, Petersen, & Brown, 2017); reduced autobiographical memory specificity (Walker, Yancu, & Skowronski, 2014); and compromised fear-learning mechanisms, such as poor safety learning, enhanced generalization of threat perception, and poor extinction learning (Gazendam, Kamphuis, & Kindt, 2013; Haaker et al., 2015). It is likely that several of these characteristics rely partly on deficits in executive function and cognitive control (Hur et al., 2019).

Both animal and human research have shown that N/NE is associated with altered neural structure and function (Depue, 2009; Hariri, 2009; Ormel et al., 2013). Recent comprehensive reviews (Hur et al., 2019; Shackman et al., 2016a, 2016b) document elevated responding to threat in several brain areas, such as the amygdala, hippocampus, insula, bed nucleus of the stria terminalis (BNST), midcingulate and orbitofrontal cortex, and periaqueductal gray. Consistent evidence further suggests that the amygdala is a coordinating brain area involved in both resting-state and threat-induced responses that are associated with elevated N/NE. On the one hand, the amygdala receives input from sensory (thalamic), contextual (hippocampal), and evaluative/regulatory (prefrontal/insular) structures that flows from the ventral areas of the amygdala to the central and dorsal/posterior parts and to the BNST. On the other hand, the latter structures (central, BNST) coordinate the behavioral (e.g., avoidance, inhibition), physiological (e.g., autonomic, neuroendocrine), and cognitive (e.g., hypervigilance, attentional bias) responses characterizing negative affective states.

Acute stressors alter amygdala functional connectivity and potentiate amygdala responses to threat (Cousijn et al., 2010; Hermans et al., 2017), and these effects are stronger in individuals with N/NE (Everaerd, Klumpers, van Wingen, Tendolkar, & Fernández, 2015), suggesting sensitization of the amygdala. Repeated acute and/or chronic stress may also lead to neural changes in other brain areas, including the hippocampus and prefrontal cortex (McEwen & Gianaros, 2011; McEwen, Nasca, & Gray, 2016). These changes might affect larger brain networks subserving executive function and cognitive control such as the frontoparietal (dorsolateral prefrontal cortex, intraparietal sulcus) and cingulo-opercular network (midcingulate cortex, anterior insula, frontal operculum; Q. Li et al., 2017). Emerging evidence suggests that frontoparietal areas such as the dorsolateral prefrontal cortex are less efficient in allocating resources to execute function tasks in individuals with N/NE (for a review, see Hur et al., 2019).

N/NE and psychopathology

A major scientific challenge remains, however, to parsimoniously conceptualize N/NE as a general transdiagnostic psychobiological dimension to solve the multifinality problem (how a transdiagnostic factor causes different disorders) and the problem of divergent trajectories (why different individuals with the same transdiagnostic factor develop different disorders; Nolen-Hoeksema & Watkins, 2011). Nolen-Hoeksema and Watkins (2011) suggested solving these problems by sorting a large number of separate processes into a coherent framework consisting of distal, proximal, and moderating (risk) factors. In another attempt at solving these problems, Barlow, Ellard, et al. (2014b) described a complex of vulnerability factors (triple-vulnerability theory) involving (a) a biological (heritable) vulnerability factor predisposing emotion-related brain structures to hyperexcitability, (b) a (learned) psychological vulnerability factor representing “a pervasive sense of unpredictability and uncontrollability in relation to life events and a perceived inability to cope with negative outcomes from such life events” (p. 484), and (c) a (learned) factor that determines why an individual becomes concerned about a particular type of threat and thus develops a particular type of disorder. The Research Domain Criteria project also tries to identify and integrate dimensional constructs at multiple levels of measurement (from genes to self-reports) in different domains of functioning (Cuthbert & Insel, 2013). For example, Lang and coworkers suggested a dimension of psychophysiologic responding in the negative-valence system that cuts across DSM-based clinical anxiety diagnoses representing diminishing defensive reactivity (or blunting of the defensive system) with decreasing focal fear and increasing negative affectivity, functional interference, and distress of the patients (Lang, McTeague, & Bradley, 2016; McTeague & Lang, 2012).

These approaches have created coherent conceptual schemas that elucidate how general threat sensitivity may lead to various psychopathological phenomena through the operation of an array of empirically established processes. However, in this article we suggest that taking a new and more radical functional perspective may deepen our understanding in a more parsimonious way. From an evolutionary perspective, N/NE as a trait can be seen as an adaptive response to repeated experiences with threat involving a recalibration of the threshold for threat detection and concomitant vigilance and hyperarousal. An optimal threshold for threat detection depends on two factors: the probability of an aversive event and the relative payoffs of the four possible outcomes of detection (i.e., true and false positives and negatives; Nettle & Bateson, 2012). Given the typically much larger costs of false negatives (missing threat) compared with false positives (metabolic expenditure), it is adaptive to recalibrate thresholds to lower levels when aversive experiences are likely, representing a BSTS strategy to cope with threat. Because of the autocorrelated nature of the person-environment system within individuals and across generations, it is not surprising that recalibration to lower thresholds has been observed in animals and humans, especially after repeated (early) adverse experiences, and that it can be transferred to subsequent generations (Baker, Cesa, Gatz, & Mellins, 1992; Hanson, Hariri, & Williamson, 2015; Hanson, Nacewicz, et al., 2015; Hariri & Holmes, 2006; Kendler, Prescott, Myers, & Neale, 2003; McEwen et al., 2016; Nettle & Bateson, 2012). Recalibration occurs in both animals and humans, suggesting a pervasive change in the way living animals construe and behave in the world that is not always captured by assessing cognitive operations and/or verbal expressions. In addition, it suggests that bias should not be considered irrational and dysfunctional but rather quite rational and adaptive in view of a person’s history (Gilbert, 1998; Nemeroff, 2013).

In line with this functional perspective, the generalized unsafety theory of stress (GUTS; Brosschot, Verkuil, & Thayer, 2016, 2017, 2018) suggests that the default state of organisms is to expect and prepare for threat and that only the perception of safety can inhibit this default state. Individuals with high trait N/NE would be unable to perceive safety, resulting in a chronic state of threat, even when none is present. Brosschot and colleagues consider different behavioral characteristics such as attentional bias, negative interpretation biases, perseverative thinking, poor safety learning, and chronic stress to represent a strategy of “erring on the safe side.” However, whereas GUTS describes why these phenomena should be interpreted this way, it does not explain how information is processed representing this strategy and how precisely this leads to the behavioral phenomena that can be observed in clinical practice.

Clarifying this “how” in terms of processing heuristics in the brain is the purpose of the current article. We suggest that interpreting N/NE and associated cognitive, affective, and behavioral phenomena within a predictive-processing perspective elucidates the centrality and communality of a BSTS strategy in a way that would not be apparent without these new conceptualizations. Predictive-processing accounts entail a radical constructivism in which the brain is seen as an organ that actively constructs a model of reality from noisy input using information it already has. Prediction signals from models in the brain are matched with sensory input, resulting in prediction errors that are fed back to improve the adaptivity of these models when making perceptual inferences and actively navigating the environment. In a broad sense, we suggest that individuals who are highly sensitive to threat tend to abort this error-reduction process in a premature phase. Although this strategy may be rewarded by beneficial effects in the short term because it helps individuals to quickly classify information about threat and negative valence, it ultimately leads to insufficient updating of their model of reality, resulting in persistent deviations between expected and actual input. We suggest that such a stagnated error-reduction process is the core of the psychobiological dimension of N/NE and that phenomenally different clinical characteristics that are associated with it are at a more fundamental level implementations of one and the same process. Before elaborating on this point, we briefly introduce predictive-processing models (for more elaborate introductions, see Barrett & Simmons, 2015; Clark, 2013; Friston, 2010, 2013; Friston, FitzGerald, Rigoli, Schwartenbeck, & Pezzulo, 2017; Hohwy, 2012, 2013; Seth, 2013; Stephan et al., 2016; Van den Bergh et al., 2017; Wiese & Metzinger, 2017).

Predictive processing, perception, and action

According to the predictive-processing framework, a basic task of the brain is to construct an adaptive model of the (external and internal) world, although its only source of information to do so is the spatial and temporal patterning of its own neural activity. Models are considered adaptive when they allow one to quickly and efficiently infer the sources of stimulation (perception) as well as to predict future states and consequences of actions (planning and action). To achieve this goal, the brain uses information not only from neural activity that is triggered by peripheral input (sense organs and receptors in the peripheral body) but also from neural activity generated by the brain itself, reflecting previous experiences and “built-in” information that act as predictions. The theory of predictive coding or processing specifies how the brain makes sense of the stimulation it receives given certain expectations using principles akin to Bayesian inference.

Input from the (external or internal) world leads to two counterflowing streams of neural activation across several hierarchical levels of the brain: Stimulation by peripheral input (“likelihood”) interacts with activations generated by the brain (“generative model”) that act as probabilistic predictions of the input (“prior beliefs”) within a specific context—that is, they estimate a likelihood of what the new input will be given previous experience. It should be noted, however, that prior beliefs are implicit assumptions of the brain that in most cases are not accessible to consciousness. Thus, they can be quite different from conscious beliefs. The discrepancy between predicted and actual inputs across multiple hierarchical levels—from low-level sensory input to high-level abstractions—results in prediction errors that are propagated throughout the system in a process of error minimization. The brain will eventually settle on/infer a posterior model that represents the most likely model of the stimulation. For example, if one is waiting for Jeff in a crowded street, the brain generates neural patterns acting as prior beliefs that will facilitate spotting Jeff in the crowd (example from Pezzulo, Maisto, Barca, & Van den Bergh, 2019, p. 3).

The system realizes error minimization in three ways: by adapting the prior beliefs to accommodate the actual input (belief update), by actively operating on the world and generating input that fits the prior beliefs (“active inference”), and by changing how the brain samples (or attends to) sensory input (Barrett & Simmons, 2015). Active inference acknowledges that perception does not passively wait for sensory input but depends on action (e.g., active sampling) to produce input. For example, waiting for Jeff may prompt the person to move toward a location that provides a better overview of the passing crowd and/or increases the scanning rate, generating more detailed information to help spot him (Pezzulo et al., 2019). In a broader sense, however, it refers to any kind of perceptual, somatovisceral, and behavioral (goal-directed) responses to produce input that is consistent with the expectations specified in the generative model to minimize prediction error. Active inference therefore involves “policy selection” to reach epistemic and/or pragmatic goals (Friston, FitzGerald, Rigoli, Schwartenbeck, & Pezzulo, 2016). This process of error minimization eventually settles on posterior beliefs that best account for the prediction errors and that reflect the combined influence of priors and the actual input. Conscious experience is thought to correspond to the posterior beliefs that are the most likely explanation of what is happening in the world (Hohwy, 2012).

An interesting consequence of this perspective is that it integrates perception, action, and physiological regulation within the same theoretical account (for a hierarchical predictive coding model describing how multiple neurovisceral interactions regulate heart rate variability in interactions with physiological and psychological demands, see R. Smith, Thayer, Khalsa, & Lane, 2017).

Precision and precision control

Priors, prediction errors, and posterior beliefs are conceived of as probability distributions that represent statistical regularities in neural activity with a mean and a variance. The inverse of the variance of these distributions is its precision. Highly precise priors and prediction errors reflect a neural pattern that has a high probability of being associated with a particular input. The opposite is true for low precise priors and prediction errors. For example, if Jeff is unusually tall, both priors and prediction errors representing Jeff’s height are highly precise, resulting in a quick and reliable recognition of Jeff. The relative impact of the distributions representing prior beliefs (i.e., predictions) versus the distributions representing the input (i.e., the likelihood of what is present) on the posterior beliefs (i.e., what the system concludes is present) will be determined by the precision of the distributions. For example, when it is dark, there is a high probability of recognizing Jeff in any tall person, reflecting a strong effect of the prior on the eventual perception. Conversely, on a sunny day it is less likely to take any tall person for Jeff, and this likelihood is even further reduced if one is not waiting for Jeff.

Because the brain cannot know whether any residual prediction error represents random information or is amenable to further minimization, it must learn the conditions under which particular models are likely to be adaptive. This is accomplished by developing context-dependent expectations about the precision of its inputs, which in turn determine how much weight is given to the prediction errors in the process (“precision optimization”; Hohwy, 2012). The implication is that contextual cues may have an important impact on the eventual posterior model that corresponds to conscious experience. In the case of Jeff, the perceptual information related to Jeff’s height is highly precise, and the brain will learn to consider height to be a highly precise prior for recognizing Jeff. In a more general way, precision control is extremely important in creating adaptive models of the stimulation that reaches the brain (Parr & Friston, 2018). Figure 1 illustrates how the relative precision of priors and prediction errors (likelihood) affects the eventual posterior model.

OPEN IN VIEWER

Fig. 1.

Integration of prior knowledge and sensory evidence (likelihood) in Bayesian inference. The top panel shows that if the prior and likelihood have the same precision (i.e., inverse variance of the Gaussian distribution), the posterior belief is in between. The middle and bottom panels show that a higher-precision prior and likelihood “attract” the posterior, respectively. Note that in all cases the precision of the posterior increases compared with the prior. Reproduced from Pezzulo, Maisto, Barca, & Van den Bergh (2019), which was published under a Creative Commons Attribution (CC BY) 4.0 International License.

Major advances have been made in relating these predictive-processing concepts to a neuroanatomical architecture and in operationalizing them in formal computational ways (see Friston, Rosch, Parr, Price, & Bowman, 2018; Parr & Friston, 2018; J. A. Roberts, Friston, & Breakspear, 2017).

Some implications

Several implications follow from this way of understanding the functionality of the brain. First, conscious experience always reflects prior expectations to some extent, but the degree to which this happens can vary. Prior beliefs with high precision in the context of imprecise inputs are likely to have a strong impact on the posterior beliefs, whereas the reverse is true when high-precision inputs are processed while prior beliefs are imprecise. In both cases, however, the same compelling sense of being “real” or “true” is produced. Second, the criterion for the brain to settle on particular posterior beliefs is not accuracy but adaptiveness or usefulness. Although accurate beliefs are mostly adaptive, it can be more adaptive to be biased (Lynn & Barrett, 2014). This suggests that it might be more fruitful not to consider bias a mistake or “being wrong” that should be corrected. Instead, trying to understand its usefulness and modifying conditions to make it less adaptive might be a more fruitful approach for modifying the information-processing “heuristics.” Third, and most importantly, predictive processing does not assume a self or an agent that organizes the information-processing traffic. Instead, the brain is considered to self-organize following the principle of free-energy minimization, and the sense of self or agency is a product of processing heuristics themselves producing perceiver and percept at the same time (Friston, 2010; Hume, 1739/2007). It follows that N/NE should not be considered to be something that a self “has” or to be some force that affects and skews “normal” information processing and should be controlled or regulated by a superseding self. Rather, we suggest that N/NE should be considered inherent to the data-processing heuristics themselves (Petersen, von Leupoldt, & Van den Bergh, 2015). The implication is that different characteristics of N/NE should not be seen as separate phenomena that are associated with N/NE but as a representation of these very data-processing heuristics. In this way, we suggest a more parsimonious interpretation of N/NE and its associated phenomena with potentially important clinical implications. We expand on this perspective below.

Perseverative cognition

Perseverative cognition, or the repetitive, sustained activation of cognitive representations of past stressful events or feared events in the future, refers to a class of cognitive activities subsuming depressive rumination and anxious worry that is both prospectively and cross-sectionally associated with anxiety disorders and depression (Brosschot, Gerin, & Thayer, 2006; Drost, Van der Does, van Hemert, Penninx, & Spinhoven, 2014). It can therefore be considered a transdiagnostic feature of internalizing disorders. Although perseverative cognition does not directly cause depression and anxiety, it acts as a mediator between N/NE and these disorders, enhancing their probability. Perseverative cognition is also associated with increased allostatic load on the cardiovascular (blood pressure, heart rate, heart rate variability) and endocrine (e.g., cortisol) systems (Ottaviani et al., 2016). Perseverative cognition is part of a broader class of repetitive thinking that includes “thinking attentively, repetitively or frequently about one’s self and one’s world” (Watkins, 2008, p. 163), but the maladaptive variant typically has a negatively valent content.

Perseverative cognition is apparently triggered by negative affective states, more precisely when such states are elicited by the awareness of a difference between the current state and a target state (J. M. Smith & Alloy, 2009). For example, depressive rumination can be seen as a cognitive elaboration of one’s current sad state and its potential negative consequences in attaining a more desired state, as well as an elaboration of the reasons for the discrepancy (“why” questions).

Perseverative cognition is characterized by an abstract level of construal (Watkins, 2008) that is described as forming general, superordinate, and decontextualized mental representations that convey the “essential gist and meaning” of events and actions, whereas concrete low-level construals include contextual, specific, and incidental details of events and actions (Watkins, 2008, p. 187). It is assumed that abstract construal is selected as an emotion-focused coping style because it allows cognitive avoidance of a thorough experience of negative affective states. This results in a short-term benefit (avoiding intense negative affect) but also a long-term problem because more adaptive processing of negative affect and active problem solving is impeded (J. M. Smith & Alloy, 2009). The overall effect is prolonged stress and negative affect, which promotes a vicious cycle with perseverative cognition as the motor that keeps it going.

The idea that perseverative cognition is a recurrent attempt to reduce discrepancies between actual and desired goals is central to control theory (Martin & Tesser, 1989, 1996) and in an extensive elaboration of it by Watkins (2008). Because both control theory and predictive processing are rooted in theories of inference and control in biological systems, it is not surprising that there are striking similarities with the current predictive-processing account of N/NE: The option to select a short-term benefit by avoiding more elaborate processing of negative content at the expense of a larger problem in the long term represents a BSTS strategy that is characterized by a stagnated error-reduction process. Rather than processing specific and contextualized events and actions that are rich in specific and concrete evidence (prediction errors) and to accommodate higher-order mental representations accordingly (error minimization), abstract construals that largely reflect prior assumptions about the individual in interaction with the world dominate mental activity, compromising adaptiveness of the mental models of oneself and the world in the long term and leading to chronic uncertainty. As noted by Watkins (2008, p. 192), this strategy is particularly disadvantageous in conditions of novelty, unfamiliarity, difficulty, or stress, that is, conditions that are typically threatening for individuals with high trait N/NE.

Although both ways of interpreting perseverative cognition are quite similar, the benefit of understanding perseverative cognition within a predictive-coding perspective is that it reveals identical mechanisms that underlie both perseverative cognition and N/NE. This suggests that perseverative cognition is not just a separate phenomenon that is associated with N/NE. Rather, perseverative cognition is at the cognitive level a very instance of the same processing style that constitutes N/NE. Put differently, it is N/NE at work when dealing with discrepancies between actual and desired goals. In this way, a predictive-processing account is more parsimonious.

Reduced autobiographical memory specificity

Reduced autobiographical memory specificity is typically assessed with the autobiographical memory test (Raes, Hermans, Williams, & Eelen, 2007; Williams & Broadbent, 1986), in which individuals are given an emotional cue word and are asked to recall specific autobiographical memories in relation to that cue word. A specific memory refers to a particular event that lasted less than a day, whereas an overgeneral memory is more generic and typically includes a class of events (categoric memories) or an extended period of time (extended memories; Griffith et al., 2009). The difficulty in retrieving specific personal memories of a past event is found in a broad range of psychopathological disorders, most notably in depression and posttraumatic stress disorder but also in acute stress disorder and somatic symptom disorder (Barry, Chiu, Raes, Ricarte, & Lau, 2018; Walentynowicz, Raes, Van Diest, & Van den Bergh, 2017). Reduced autobiographical memory specificity is also a marker of an unfavorable course of psychopathology that affects the severity of symptoms, illness duration, and treatment success (Sumner, Griffith, & Mineka, 2010); in a community sample, it predicted a gradual increase in depressive symptoms over the course of 18 months (Van Daele, Griffith, Van den Bergh, & Hermans, 2014).

Reduced autobiographical memory specificity is associated with impaired executive function (e.g., deficits in inhibitory control, updating and maintaining information in working memory, verbal fluency) and with avoidance of negative affect (Barry et al., 2018; Sumner et al., 2014). Several areas of the brain involved in processing emotional salience and self-relevance as well as in executive control, emotion regulation, and memory have been associated with reduced autobiographical memory specificity, and these effects may be due in part to abnormalities, chronic or otherwise, in cortisol. The activation patterns in brain areas are not systematically replicated across diagnostic groups with the same behavioral effects, suggesting that reduced autobiographical memory specificity may result from a number of different neurocognitive mechanisms (Barry et al., 2018). A large-scale neural-network analysis might extend and further clarify how reduced autobiographical memory specificity is implemented in the brain.

One of the important factors that is assumed to account for reduced autobiographical memory specificity is capture and rumination (CaR-FA-X model; Williams et al., 2007), which implies that memory retrieval is “captured” at a general level, thereby blocking the retrieval of specific memories, whereas ruminating keeps processing of the information at an analytical and abstract level (see above). These deficits obviously make adaptive processing of negative information less likely, leaving one stuck at the level of overgeneral, self-related information. The choice of this level of processing is assumed to result from functional avoidance of episodes of negative affect that might be triggered by emotional cues. In the long term, however, this strategy is maladaptive because it inhibits emotional processing and active problem solving related to the sources of the negative affect. An unresolved issue is whether this strategy is already in play at the encoding phase, resulting in poor specific and detailed memories of negatively valent or threatening information in the first place (Raes, 2005).

From a predictive-processing perspective, this way of processing information in memory is consistent with the BSTS strategy described above: Little detailed processing of memory input from autobiographical events results in low-precision prediction errors, allowing strong prior beliefs representing (negatively valent) generalized memory information to determine the conscious recall of self-referential events. Consequently, error-minimization stagnates at a rather abstract level, leaving generalized prior beliefs unchanged and contributing to a vicious cycle of negative affect and low autobiographical memory specificity. Again, it is clear that reduced autobiographical memory specificity, when viewed in this way, is not just a phenomenon that is associated with N/NE. It is actually the core of N/NE at work when processing self-relevant memory information.

Compromised fear learning

Associative fear learning typically involves repeated pairing of a (relatively) neutral stimulus with an aversive fear-inducing stimulus (unconditioned stimulus), after which the previously neutral but now conditioned stimulus (CS) elicits fear by itself. Extinction learning implies repeated unreinforced exposures to the CS, leading to reduced responding to it. Associative fear learning is widely used as a laboratory model for understanding pathological fear (Craske, Hermans, & Vervliet, 2018). Although strong conclusions about the effect of individual-difference variables are hampered by the “noise” induced by different methodologies, some evidence suggests specificities in associative fear learning that are associated with N/NE and the vulnerability to develop and/or suffer from actual pathological anxiety (Lonsdorf & Merz, 2017). First, anxiety-prone individuals tend to show poor extinction learning, that is, a less steep decline of fear responses during extinction and/or a higher level of remaining fear after extinction (Shechner, Hong, Britton, Pine, & Fox, 2014). Put differently, for an equal amount of expectancy violations, the CS remains categorized as potentially dangerous for a longer period of time. Poor extinction learning in the laboratory has been shown to predict the development of anxiety symptoms in real-life situations (Guthrie & Bryant, 2006).

Second, anxiety-prone individuals show impaired safety learning in a differential-learning paradigm that involves the presentation of two neutral stimuli, one of which is paired with the aversive stimulus (CS+), and the other is never paired with it (CS−), turning the neutral stimuli into, respectively, a danger cue and safety cue. Impaired safety learning, emerging as elevated fear responding to the CS−, has frequently been observed in individuals scoring high on trait anxiety (Gazendam et al., 2013; Kindt & Soeter, 2014), in individuals with subclinical levels of anxiety (Arnaudova et al., 2013; Chan & Lovibond, 1996; Haddad, Pritchett, Lissek, & Lau, 2012), and in individuals with anxiety disorders (Grillon & Morgan, 1999; Jovanovic et al., 2013; Lissek et al., 2009; Winslow, Noble, & Davis, 2008) and depression (Pollak et al., 2008). Other data have shown that patients with panic disorder had higher danger expectancy when safety stimuli were presented (Lenaert, Boddez, Vervliet, Schruers, & Hermans, 2015). Poor danger-safety discrimination also predicts the return of fear after treatment (Staples-Bradley, Treanor, & Craske, 2018). In other words, individuals who are vulnerable to or with affective psychopathology show reduced discrimination between danger and safety in a danger context, illustrating a bias toward categorizing a safe cue as dangerous (see also Garcia & Zoellner, 2017).

Third, overgeneralization of fear has also been observed in threat-sensitive individuals. Generalization occurs when fear has been learned in response to one stimulus and subsequently emerges in response to stimuli that are similar but have never been paired with the aversive stimulus. Overgeneralization, then, occurs when fear responding to generalization stimuli remains relatively higher than healthy controls across a decreasing gradient of similarity with the original CS. This suggests that threat-sensitive individuals are less sensitive to differences between a new stimulus and danger cues, leading to a higher probability of categorizing new stimuli as dangerous. Overgeneralization of learned fear has been found in several groups with anxiety disorders (Lissek et al., 2010, 2014; Lissek & Grillon, 2012; Lenaert et al., 2014).

Fourth, a general sense of unpredictability and uncontrollability contributes to the development of neuroticism. This may result from a lack of safety learning and/or repeated experiences with unpredictable and uncontrollable aversive life events and from harsh, intrusive, and overcontrolling parenting styles. It is associated with alterations in neural function and neuroendocrine regulation of the hypothalamic-pituitary-adrenal axis (see Barlow, Ellard, et al., 2014).

In sum, individuals prone to developing emotional disorders are less likely to develop inhibitory fear learning, as indicated by poor extinction, poor safety learning, and overgeneralization, and they experience the world as less predictable and controllable. Put simply, they are more likely to categorize a safe stimulus as potentially dangerous, need more expectancy violations to shift a danger cue into a category of safe ones, and—more so than less threat-sensitive individuals—tend to categorize as dangerous both a safe cue within a dangerous context as well as a new cue that resembles a dangerous one.

How do these effects come about according to a predictive-processing framework? As suggested above, these effects may result from applying a perceptual decision rule that has shifted toward oversimplifying input (impoverished sampling), allowing highly precise threat-related priors to have a stronger impact on the eventual categorical danger perception. The obvious benefit is a higher likelihood to categorize input as threat at the expense of the level of detail by which the input is processed. This can be considered an instance of ecologically situated perception (Linson & Friston, 2019). Evidence consistent with this interpretation has been found in relation to fear overgeneralization in individuals with anxiety. Rather than resulting from postperceptual-choice behavior, it reflects errors resulting from an altered perception of simple stimulus features and causes less discrimination between danger and safety cues (Laufer, Israeli, & Paz, 2016; see also Struyf, Zaman, Vervliet, & Van Diest, 2015; Zaman, Ceulemans, Hermans, & Beckers, 2019). Other evidence in support of this interpretation may be found in studies linking these learning abnormalities to memory specificity (see above). Accurate discrimination learning relies on sufficiently detailed and specific memory: Without detailed memory, experiences will not be represented in memory as unique, specific events, reducing the ability to differentiate danger from safety. Lenaert et al. (2015) showed that individuals with anxiety who had low memory specificity also showed poor discrimination learning in a differential-conditioning paradigm. It may also be relevant in this context that animal studies have shown that stress-related changes in the hippocampal area lead to deficient “pattern separation” during memory encoding, making distinct patterns of information (e.g., CS+ vs. CS−) less distinguishable in memory (Tronel et al., 2012).

The overall picture of these learning abnormalities suggests that individuals with high trait N/NE use an abstract, general way of construing sensory input, leaving conscious experience and behavior relatively more influenced by categorical threat-related priors than by actual sensory evidence. Again, this is compatible with a BSTS strategy when processing aversive information: Stimuli are more likely classified as aversive or dangerous at the costs of more errors “on the safe side.” This interpretation again suggests that N/NE-related learning abnormalities are not just phenomena associated with N/NE. They actually are the N/NE-related processing heuristics in action during fear learning that lead to altered phenomenal learning characteristics.

Symptom perception in individuals with somatization and medically unexplained symptoms

A pervasive positive relationship has been established between trait N/NE and symptom reports unrelated to physiological dysfunction, or so-called medically unexplained symptoms (MUS; Van den Bergh et al., 2017). Such symptoms are highly prevalent throughout the health-care system (Haller, Cramer, Lauche, & Dobos, 2015): They appear in individuals who are not consulting health care providers, such as those with high trait N/NE, who tend to have more symptoms in daily life; in primary care, in which patients with MUS are characterized by elevated levels of anxiety and depression; and in secondary care, in which functional somatic disorders (or somatic symptom disorder) are associated with elevated psychiatric comorbidity (Henningsen, Zipfel, & Herzog, 2007; Wessely, Nimnuan, & Sharpe, 1999; Witthöft & Hiller, 2010).

A clue to understanding MUS may be found in the fact that neurobiological and psychometric evidence (Walentynowicz, Witthöft, Raes, Van Diest, & Van den Bergh, 2018) suggest that the experience of a symptom entails a sensory-perceptual component and an affective-motivational component that are intuitively integrated into one symptom report. Individuals with high trait N/NE show enhanced affective-motivational responding to aversive somatic sensations as well as reduced detail in sensory-perceptual processing, blurring a clear distinction between an emotional and a somatic state (Van den Bergh & Walentynowicz, 2016). This results in little precise somatic input (prediction errors), allowing for a large impact of symptom-related categorical priors that are constantly activated in these individuals. In other words, in individuals with MUS, categorical perception of the somatic state is largely determined by somatic priors and less or not at all by distinct sensory input (Van den Bergh et al., 2017).

Less detailed sensory-perceptual processing of somatic episodes is suggested by a number of recent findings:

Other evidence is compatible with the idea that input from somatic and affective states is not very well distinguished in individuals with MUS, making them highly vulnerable to somatic priors when in a negative affective state. For example, when inducing negative affect through picture viewing followed by a symptom questionnaire that activates symptom-related priors, elevated symptom reports are elicited in these individuals, and this effect is mediated by the difficulty of identifying feelings (Van Den Houte et al., 2017). Moreover, harmless cues that have been associated with symptom episodes in a conditioning paradigm can elicit elevated symptoms by themselves. This happens particularly when the cues have a negative affective valence and in individuals with high trait N/NE and individuals with somatic symptom disorder (Devriese et al., 2000; Van den Bergh, Stegen, & Van de Woestijne, 1997).

In sum, the above pattern of results converges on the interpretation that individuals with high trait N/NE are more affected by the aversive affective-motivational component of a somatic episode and process sensory-perceptual information of a somatic episode in a less detailed way. This results in fewer precise prediction errors and leads to persistently active symptom-related priors that produce symptoms that are poorly related or unrelated to physiological dysfunction. We suggest that this is an instance of a BSTS strategy: Despite little sensory evidence, the brain jumps to the conclusion of an aversive somatic symptom being present. The benefit is a reduction of uncertainty at an early stage of the type “better the devil you know than the one you don’t” (Van den Bergh et al., 2017, p. 197). The downside, however, is little sensitivity to error correction by means of updating the prior expectation, which leads to persistent physical symptoms unrelated to actual physiological dysfunction (Henningsen et al., 2018). Again, MUS are in this respect not a separate phenomenon that is associated with high trait N/NE. It actually is the N/NE-related processing heuristics at work when processing somatic information.