Intro: Attentional Networks in BPD
Affective instability (AI) is a hallmark symptom of Borderline Personality Disorder (BPD). Although affective variation is common across pathologies of mood, the rapidity and intensity of affective shifts in BPD are striking and may be distinct from other mental disorders. In the clinical neuroscience literature, AI has been associated with fronto-limbic dysfunction, particularly the reciprocal connections between the medial prefrontal cortex (mPFC)/ anterior cingulate cortex (ACC) and the amygdala (Dudas et al., 2017; Minzenberg, Fan, New, Tang, & Siever, 2007; New et al., 2007; Schulze, Schmahl, & Niedtfeld, 2016).
Recent research, however, has failed to demonstrate that altered PFC-amygdala interactions are sufficient to produce the BPD clinical phenotype, especially AI. Further, such studies of “emotional reactivity” in BPD often rely on viewing static emotional pictures and faces, which are relatively non-specific indices of emotional “reactivity” that do not speak to the effects of emotional cues on behavior. In short, static face paradigms provide only rudimentary information about how individuals with BPD respond to negative or threatening situations. Complicating these interpretations further, recent results from resting-state functional connectivity studies suggest that altered salience processing may play an important role in BPD. Likewise, we have suggested that circuits supporting value-based decision-making may be a key component of the neurobiological basis of BPD (Doll et al., 2013; Hall and Hallquist, in prep; Sarkheil, Ibrahim, Schneider, Mathiak, & Klasen, 2019). Altogether, recent neuroscience studies of BPD highlight the anterior insula (aIns), ventral striatum (VS), ventromedial prefrontal cortex (vmPFC), and temporoparietal junction (TPJ) as key targets for further investigation in BPD.
Of particular relevance to the current proposal, recent findings from our group suggest that the right aIns shows robust patterns of hyperconnectivity in a sample of adolescents and young adults with BPD compared to healthy controls (Hall & Hallquist, in prep). The right aIns, in concert with the dACC form the core of the putative salience/ventral attention network (VAN; Corbetta & Shulman, 2002). The insula likely mediates shifts in attention based on multimodal input, ranging from sensory and interoceptive signals to higher order cognitive/representational signals from the cortex (Menon & Uddin, 2010; Sridharan, Levitin, & Menon, 2008; Uddin, 2015). The ventral attention network, most notably the right aIns is also thought to control the relative balance of internal vs externally oriented attention by acting as a “circuit-break” to ongoing processes when motivationally salient information needs to be attended to in a bottom-up fashion (Corbetta & Shulman, 2002; Wang et al., 2015).
In our resting-state data with BPD participants we found that the association between group status (BPD/Control) and heightened connectivity in the right aIns was statistically mediated by self-reported levels of affective instability. This is consistent with a tentative hypothesis that heightened intrinsic connectivity of the aIns corresponds to a heightened propensity for attentional switching in BPD, which may lead to frequent shifts in emotional experience. For example, connectivity of aIns may support sensitivity to the internal state, diverting attention from a goal at hand (e.g., grocery shopping) to emotionally congruent cues (e.g., an ex-girlfriend the grocvery aisle).
The aIns is involved in signaling prediction errors associated with punishment (Harrison et al., 2016; Pessiglione, Seymour, Flandin, Dolan, & Frith, 2006). However, alternative accounts suggest that the insula is implicated in signaling the salience of stimuli more generally, including reward cues. This has led to a hypothesis that the aIns signals the absolute value of a prediction error or “salience prediction error” (Metereau & Dreher, 2013), which can be utilized to determine the need to execute a motivationally-relevant behavior or rely on self-reflective circuitry in the default mode network (DMN; Sridharan et al., 2008).
Important to the current proposal, in our resting-state data, we also found robust hypo-“activation” (measured via the amplitude of low-frequency fluctuations; ALFF) in the Dorsal Attention Network (DAN) in our BPD participants (this hypoactivation also decreased with age). This tentatively suggests that the “push-pull” dynamics between VAN and DAN previously described in healthy participants (Corbetta & Shulman, 2002; Vossel, Geng, & Fink, 2014) favors VAN in individuals with BPD. These findings imply that the more goal-directed or “phasic” control of attention (which is largely regulated by DAN) may be compromised in individuals with BPD.
Attention and Reinforcement Learning in Healthy Participants
Recent evidence from cognitive neuroscience shows that attention narrows/reduces the dimensionality of rich sensory information that is relayed for further processing, perhaps via precision-weighted expectations, as described in theories of predictive coding (Spratling, 2008). This can be implemented by systems that exert either top-down (DAN) or bottom-up (VAN) influence (i.e. biasing) on the visual system. Recently, this biasing of attention has been formally incorporated into a computational reinforcement learning framework. In particular one recent study demonstrated that attention aids reinforcement learning by biasing eye gaze towards the features of compound stimuli that contain information that is the most predictive of rewards (Leong, Radulescu, Daniel, DeWoskin, & Niv, 2017). The authors found that greater attention during decision-making biased choice, such that highly attended items were biased towards being chosen. Further, behavior on this task was best fit by augmenting standard delta rule learning with a parameter that weighted the magnitude of a prediction error on stimulus representation (e.g., face) according to attentional allocation to a stimulus during the trial. The reciprocal relationship between attention and learning systems can help researchers understand how humans successfully implement the relatively simple rules of reinforcement learning in complex/multidimensional environments.
The Current Study
Stepping back to the level of clinical description, under acute stress, patients with BPD are prone to exhibit a number of maladaptive decisions including impulsive behaviors (self-harming behaviors, brief sexual encounters with strangers, binge drinking, etc.) and/or relational aggression (yelling at a partner, hitting a partner, crying for fear of a partner leaving, etc.). One largely unexplored area of interest in BPD research is whether under stress, attentional systems that often help reinforcement learning circuits function efficiently (by reducing the complexity of the environment) play a role in promoting maladaptive decisions. For example, during an argument with a significant other, persisting with the effortful goal of working through the disagreement might be supplanted by excessive attention to the angry expression on a partner’s face (cf. Miano, Dziobek, & Roepke, 2017). This “emotional hijacking” of attention may promote emotion-congruent behaviors such as responding defensively, yelling, or developing an acute concern that the relationship is coming to an end. While the above example draws on clinical experience and theory in patients with personality pathology, it is likely that increased attention to emotionally congruent stimuli under stress occurs in the normative population. Under stress, it is likely that attentional resources are “hijacked” from goal-directed learning systems and diverted to systems that orient agents to motivationally relevant cues. Furthermore, both acute and chronic stress may shift the balance from goal-directed toward Pavlovian and habitual actions (e.g., as reviewed in Arnsten et al., 2009).
I predict that this potential “hijacking” of attention by acute stress is mediated by an increased influence of VAN relative to DAN. Further, I expect that the influence of the VAN during stress is determined in part by the level of estimated uncertainty about the environment. In a more complex scene, the dispersion of visual attention to different stimuli may provide a useful index of estimated uncertainty (where should I look?) and may be quantifiable by the entropy of eye gaze across the 2D plane or among specific areas of interest.
Basic work on the interplay between uncertainty and the sympathetic nervous system have shown that pupil diameter and SCRs are both tightly related to arousal and scale with “irreducible uncertainty” (i.e. uncertainty that is inherent to the task; de Berker et al., 2016; Nassar et al., 2012). This suggests that when an agent is in a high entropy state (e.g.,the values of alternative actions, or which cue is associated with an outcome, is roughly equal across available actions), physiological indices of arousal may be be associated with a greater need to explore the environment or update one’s beliefs. As mentioned above, attention plays a crucial role of guiding agents through multidimensional environments. Thus, when an agent feels threatened it may, regardless of irreducible uncertainty, increase its representation of uncertainty about the current state in order to remain vigilant to potential cues or actions that could lead to punishments or rewards. This increased vigilance may be reflected in the overall dispersion of attention (perhaps as measured by higher entropy in the gaze time distribution across a scene), which likely aids in threat detection and promotes attentional shifts (via the VAN), perhaps at the cost of engaging in goal-directed attention (via DAN).
Importantly, the role of attentional control on behavior in BPD has not been investigated. Using two learning-based experimental tasks combined with fMRI, eye-tracking, and computational modeling of behavior, this proposal attempts to be the first exploration into the role of stress and attention in shaping value-based decision-making in BPD.
Study 1
In the first study, we plan to extend the multi-feature RL task implemented in Leong, Radulescu, et. al. (2017) to include blocks of trials in which we experimentally manipulate arousal with the sustained threat of an electrical shock (figure 1). Threat of shock paradigms have unique advantages over other emotion induction paradigms currently used in the BPD literature, including a relatively simple design, few task purity-related concerns (for example, the mechanism of action in standard emotional faces paradigms is vague at best), and a good track record of producing transient and prolonged increases in sympathetic nervous system activity, reflected by both skin conductance responses (SCRs; Kopacz & Smith, 1971; Robinson, Vytal, Cornwell, & Grillon, 2013) and pupil diameter (de Berker et al., 2016; Leuchs, Schneider, Czisch, & Spoormaker, 2017). Further, threat of shock paradigms have long held an important place in neurophysiological research on adaptive and maladaptive anxiety, where the goal is to excamine the anticipation of uncertaint shocks rather than responses to shock itself. The goal of the first study will be to examine how the emotional arousal elicited by the threat of shock impairs decision-making.
Study 1 Preliminary Predictions
The primary prediction is that relative to the control group, individuals with BPD will exhibit impaired learning of which stimulus dimension best predicts rewards due to rapid shifting among stimulus dimensions (people, houses, tools) in a more stochastic manner, which will account for group differences in performance. More specifically, I predict that the threat condition will lead to more exploration of the feature space for all participants (in the figure, notice the proportion of gaze in the top right panel increases for the two incorrect dimensions, In1 and In2). But I further hypothesize that this effect will be particularly pronounced in the BPD group.
This group difference in the effect of threat on attentional biasing and performance will be accounted for by group differences in entropy, either of the learned feature values or of the gaze time distribution across stimuli. (Indeed, perhaps by attending broadly, the learned feature values will also be more similar – i.e., diffuse attention promotes diffuse credit assignment.) Further, we hypothesize that impaired learning during threat in BPD participants will partly reflect their failing to reduce entropy later in learning compared to their control counterparts (bottom right) – that is, BPD participants will fail to learn to home in on the rewarded stimulus dimension.
Task 1 alternatives, design considerations
We need to consider whether the outcome domain – gains versus losses – needs to be compatible with the threat vs. safe manipulation. For example, will a threat of shock + instrumental learning of rewards lead to blunted/slowed learning, as in the de Berker 2016 paper where stress impaired learning to act? Conversely, would shock promote learning to avoid losses in the game? It could benefit us to include punishments in this task, but it’s also another ‘moving part’ in the task.
Michael had a comment that we may want to manipulate stimulus salience specifically given the VAN hypotheses. For example, we could manipulate visual salience of stimuli independent of their reward value. One option would be to manipulate the luminance of stimuli in a counterbalanced fashion to inject task-irrelevant salience into the stimuli. This would give the experiment a bit more precision to test the idea that attention isn’t merely more diffuse under threat (like scanning the periphery for predators), but that attention goes toward visually salient stimuli regardless of their value (relating to attentional capture).
Study 2
I want to preface that this portion of the study is less formed. The intuition is there, though we have not had the chance to work through all of the relevant considerations (e.g. in the alternatives, considerations section).
We have recently proposed that Pavlovian influences likely dominate goal-directed decision-making in individuals with BPD, especially in emotionally arousing contexts (Hallquist, Hall, Schreiber, & Dombrovski, 2018). In the context of social interactions, this may take the form of bringing task-irrelevant prior information to bear on one’s decisions in the interaction. For example, if I have a strong prior that others are threatening or harmful, I may struggle to develop a sense of safety in a new social encounter, even if I have no statistics on the individual with whom I am interacting. That is, we assert that individuals with BPD struggle with the ability to integrate experience that contradicts their prior.
Study 2 aims to combine our Pavlovian and attentional hypotheses. We propose to implement a novel task called something like the emotion-identity interference task. In the conditioning phase, participants will learn to associate unemotional (neutral) male and female faces from a standardized database with deterministic wins and losses or with no outcome. This phase serves to condition participants on the expected value of faces, prior to the instrumental choice task. Pupil diameter will be collected to investigate the associability of conditioned stimuli, which we believe will play an important role in biasing attention in the (instrumental) learning phase.
In the (instrumental) learning phase, subjects are shown three [MH: will there be more than three during conditioning. If so, why? NH: Haven’t thought deeply about this, but my knee-jerk reaction is to just have the three. Seems worth discussing though!] of the faces presented in the conditioning phase arranged in a single row in the center of the screen. In the learning phase, however, each face will be displayed as either happy, sad, or angry. Importantly, the emotion of the face is irrelevant to the contingency. The subjects will be told to learn how to earn as much as possible but are not told what aspect of the stimulus is most predictive of the reward. Under the hood, the payoff will follow a probabilistic three-armed bandit task with nonstationary probabilities. The subject must learn about which facial “identity” (i.e. which person, regardless of emotion) is related to the highest probability of reward. Subjects will be expected to learn an underlying reward contingency based off of the identity of the face. Provisionally, we will initialize the action-outcome reward probabilities according to the conditioning phase probabilities, but these will then shift such that subjects must learn an action-outcome contingency that does not relate to the earlier Pavlovian value gradient. (See Dombrovski, Hallquist, Brown, Wilson, & Szanto, 2019, for what such a contingency might look like.)
Thus, throughout learning there are nine stimulus classes that subjects will navigate, which are included in the matrix below:
| Sad | Happy | Angry | |
|---|---|---|---|
| CS+ | CS+, S | CS+, H | CS+, A |
| CS- | CS-, S | CS-, H | CS-, A |
| UCS (neutral) | UCS, S | UCS, H | UCS, A |
On a given trial, all three emotions and all three faces are displayed in a counterbalanced manner. Subjects are reinforced based on the underlying contingencies, which are assigned to face identities, rather than emotions. As described above, initially the contingencies are specified such that the Pavlovian CS+ is reinforced on close to 80% of instrumental face choices. However, subjects should learn to ignore the valence of the face as it contains no information on the underlying reward probabilities.
Study 2 Preliminary Predictions
First, we predict that both groups will learn the initial contingencies relatively quickly, which will be strengthed by the average magnitude of pupillary response in the conditioning phase. However, on trials where the rewarded stimulus’ emotion is Pavlovian incongruent, (CS+S, CS+A), the BPD group will show more evidence of exploratory gazes (i.e gazes to the UCS and CS- faces), perhaps reflective of emotional interference of the facial expressions. This may reflect a brittle Pavlovian association to cues predicting reward. Most importantly, we expect a failure to adapt to changes in the reward contingencies when the CS- becomes associated with higher chances of success.
Altogether, this variant begins to consider how attention may be shaped by previous Pavlovian association, a kind of attentional biasing of PIT, potentially.
Task 2 alternatives, design considerations
A simpler version of the task could be implemented without conflicting emotional expressions. This is the part of the task that conceptually needs to be worked through the most. Do the emotions essentially undo the learned Pavlovian values from the conditioning phase? Even if we think irrelevant emotion will hamper learning more in BPD, it may undermine the task to have participants trying to a) learn a new instrumental payoff, b) overcome the learned Pavlovian association, and c) figure out whether emotional valence of the face has predictive value.
While we’re already including a shock version of the first task, it may be fruitful to consider if threat strengthens Pavlovian predominance in our BPD individuals. We may find that individuals in the BPD group never even explore the CS- identity in a threatening context…
We have talked about a variant of the task where the payoff shifts mid-task from identity to emotion. This would be a kind of extra-dimensional shift, somewhat related to the Radulescu, Daniel, & Niv (2016) paper. This significantly complicates the task in addition to our predictions/interpretation but could provide an interesting opportunity to examine difficulty differences in the stimulus dimension (e.g., we would probably expect learning from face identity with irrelevant emotions to be harder than the converse), and whether individuals with BPD struggle to remap their value representation when emotion goes from having predictive value to being irrelevant.
References
Corbetta, M., & Shulman, G. L. (2002). Control of goal-directed and stimulus-driven attention in the brain. Nature Reviews. Neuroscience, 3(3), 201–215. https://doi.org/10.1038/nrn755
de Berker, A. O., Rutledge, R. B., Mathys, C., Marshall, L., Cross, G. F., Dolan, R. J., & Bestmann, S. (2016). Computations of uncertainty mediate acute stress responses in humans. Nature Communications, 7. https://doi.org/10.1038/ncomms10996
Doll, A., Sorg, C., Manoliu, A., Meng, C., Wöller, A., Förstl, H., … Riedl, V. (2013). Shifted intrinsic connectivity of central executive and salience network in borderline personality disorder. Frontiers in Human Neuroscience, 7. https://doi.org/10.3389/fnhum.2013.00727
Dombrovski, A. Y., Hallquist, M. N., Brown, V. M., Wilson, J., & Szanto, K. (2019). Value-Based Choice, Contingency Learning, and Suicidal Behavior in Mid- and Late-Life Depression. Biological Psychiatry, 85(6), 506–516. https://doi.org/10.1016/j.biopsych.2018.10.006
Dudas, R. B., Mole, T. B., Morris, L. S., Denman, C., Hill, E., Szalma, B., … Voon, V. (2017). Amygdala and dlPFC abnormalities, with aberrant connectivity and habituation in response to emotional stimuli in females with BPD. Journal of Affective Disorders, 208, 460–466. https://doi.org/10.1016/j.jad.2016.10.043
Hall, N. T., & Hallquist, M. N., (in prep). Developmental changes in resting-state functional connectivity in borderline personality disorder: a network analysis approach.
Hallquist, M. N., Hall, N. T., Schreiber, A. M., & Dombrovski, A. Y. (2018). Interpersonal dysfunction in borderline personality: A decision neuroscience perspective. Current Opinion in Psychology, 21, 94–104. https://doi.org/10.1016/j.copsyc.2017.09.011
Harrison, N. A., Voon, V., Cercignani, M., Cooper, E. A., Pessiglione, M., & Critchley, H. D. (2016). A Neurocomputational Account of How Inflammation Enhances Sensitivity to Punishments Versus Rewards. Biological Psychiatry, 80(1), 73–81. https://doi.org/10.1016/j.biopsych.2015.07.018
Kopacz, F. M., & Smith, B. D. (1971). Sex Differences in Skin Conductance Measures as a Function of Shock Threat. Psychophysiology, 8(3), 293–303. https://doi.org/10.1111/j.1469-8986.1971.tb00459.x
Leong, Y. C., Radulescu, A., Daniel, R., DeWoskin, V., & Niv, Y. (2017). Dynamic Interaction between Reinforcement Learning and Attention in Multidimensional Environments. Neuron, 93(2), 451–463. https://doi.org/10.1016/j.neuron.2016.12.040
Leuchs, L., Schneider, M., Czisch, M., & Spoormaker, V. I. (2017). Neural correlates of pupil dilation during human fear learning. NeuroImage, 147, 186–197. https://doi.org/10.1016/j.neuroimage.2016.11.072
Menon, V., & Uddin, L. Q. (2010). Saliency, switching, attention and control: A network model of insula function. Brain Structure and Function, 214(5–6), 655–667. https://doi.org/10.1007/s00429-010-0262-0
Miano, A., Dziobek, I., & Roepke, S. (2017). Understanding interpersonal dysfunction in borderline personality disorder: A naturalistic dyadic study reveals absence of relationship-protective empathic inaccuracy. Clinical Psychological Science, 5(2), 355–366. https://doi.org/10.1177/2167702616683505
Minzenberg, M. J., Fan, J., New, A. S., Tang, C. Y., & Siever, L. J. (2007). Fronto-limbic dysfunction in response to facial emotion in borderline personality disorder: An event-related fMRI study. Psychiatry Research: Neuroimaging, 155(3), 231–243. https://doi.org/10.1016/j.pscychresns.2007.03.006
Nassar, M. R., Rumsey, K. M., Wilson, R. C., Parikh, K., Heasly, B., & Gold, J. I. (2012). Rational regulation of learning dynamics by pupil-linked arousal systems. Nature Neuroscience, 15(7), 1040–1046. https://doi.org/10.1038/nn.3130
New, A. S., Hazlett, E. A., Buchsbaum, M. S., Goodman, M., Mitelman, S. A., Newmark, R., … Siever, L. J. (2007). Amygdala–Prefrontal Disconnection in Borderline Personality Disorder. Neuropsychopharmacology, 32(7), 1629–1640. https://doi.org/10.1038/sj.npp.1301283
Pessiglione, M., Seymour, B., Flandin, G., Dolan, R. J., & Frith, C. D. (2006). Dopamine-dependent prediction errors underpin reward-seeking behaviour in humans. Nature, 442(7106), 1042. https://doi.org/10.1038/nature05051
Robinson, O. J., Vytal, K., Cornwell, B. R., & Grillon, C. (2013). The impact of anxiety upon cognition: Perspectives from human threat of shock studies. Frontiers in Human Neuroscience, 7. https://doi.org/10.3389/fnhum.2013.00203
Sarkheil, P., Ibrahim, C. N., Schneider, F., Mathiak, K., & Klasen, M. (2019). Aberrant functional connectivity profiles of brain regions associated with salience and reward processing in female patients with borderline personality disorder. Brain Imaging and Behavior. https://doi.org/10.1007/s11682-019-00065-z
Schulze, L., Schmahl, C., & Niedtfeld, I. (2016). Neural Correlates of Disturbed Emotion Processing in Borderline Personality Disorder: A Multimodal Meta-Analysis. Biological Psychiatry, 79(2), 97–106. https://doi.org/10.1016/j.biopsych.2015.03.027
Spratling, M. W. (2008). Predictive coding as a model of biased competition in visual attention. Vision Research, 48(12), 1391–1408. https://doi.org/10.1016/j.visres.2008.03.009
Sridharan, D., Levitin, D. J., & Menon, V. (2008). A critical role for the right fronto-insular cortex in switching between central-executive and default-mode networks. Proceedings of the National Academy of Sciences, 105(34), 12569–12574. https://doi.org/10.1073/pnas.0800005105
Uddin, L. Q. (2015). Salience processing and insular cortical function and dysfunction. Nature Reviews Neuroscience, 16(1), 55–61. https://doi.org/10.1038/nrn3857
Vossel, S., Geng, J. J., & Fink, G. R. (2014). Dorsal and Ventral Attention Systems: Distinct Neural Circuits but Collaborative Roles. The Neuroscientist, 20(2), 150–159. https://doi.org/10.1177/1073858413494269
Wang, L., Yu, H., Hu, J., Theeuwes, J., Gong, X., Xiang, Y., … Zhou, X. (2015). Reward breaks through center-surround inhibition via anterior insula. Human Brain Mapping, 36(12), 5233–5251. https://doi.org/10.1002/hbm.23004