Recent years have seen unprecedented increases in compulsive behaviors, even in previously undiagnosed populations. Yet, despite their rising prevalence and the fact that compulsions—notably associated with obsessive-compulsive disorder (OCD), but also a key feature of other disorders such a hoarding or body dysmorphic disorder—significantly hamper daily function, the specific mechanisms of their formation remain unknown. Previous research holds mixed findings on compulsivity-linked cognitive deficits, and, while many accounts assume that compulsions aim to reduce anxiety, most individuals with anxiety do not develop compulsions—suggesting that the anxiety-reduction account is incomplete. We propose here a broader, uncertainty-resolving account of compulsivity, that bypasses the commonly-used, single-goal structure of many cognitive tasks to leverage the real-world hierarchical structure of goals—with “local”, short-term subgoals (e.g., “how do I clean my hands?”) in the service of long-term, “global” goals (“how do I stay healthy?”). In that framework, we suggest that compulsivity may reflect a deficit in integrating uncertainty from local to global goals, which still spares the basic circuitry of reward and learning. We tested 20 OCD patients and 20 healthy volunteers in a predictive inference task with a hierarchical structure of local uncertainty reduction in the service of global uncertainty reduction. Both groups learned the reward structure of the local environment based on observed data; however, only the healthy volunteers showed evidence of integrating learned knowledge at the local level to reduce uncertainty about the higher level of the task hierarchy. This indicates a potential mechanism for compulsions that relies on the inability to adaptively integrate, prioritize, and “toggle” among different local goals in the service of global goals.
The exploration-exploitation trade-off is ubiquitous in our everyday lives, and individuals display considerable variability in their preferred decision-making strategies. Most previous work pertaining to neural signatures of exploration is restricted to functional pathways. However, the specific contributions of cortical microarchitectures to high-level cognitive processes such as decision-making are as yet unknown. Here, we investigated the neuroanatomical foundations of inter-individual variability in decision-making strategies. To this end, 122 healthy participants completed a gamified multi-armed bandit paradigm aimed at teasing apart distinct exploration-exploitation decision strategies. We also collected whole-brain quantitative MRI maps indexing microstructural features of cortical myelination and iron content. Through computational modelling, we disentangled individual-specific exploration strategies, including value-free random exploration. Whole-brain regression analyses identified significant associations between value-free exploration and increased cortical myelination in right frontal brain areas with reported links to impulsivity. By elucidating the brain microstructural correlates of distinct exploration-exploitation strategies, we aimed to further our understanding of why individuals differ in their decision-making capabilities, and how decision-making may become aberrant in mental health conditions.
Reward valuation and reward and punishment learning are key constructs relevant for understanding and intervening on mental illness. These constructs are frequently measured with behavioral tasks such as intertemporal choice tasks, measuring delay discounting, and the Iowa Gambling Task (IGT), measuring reward and punishment learning. Delay discounting refers to how people value delayed outcomes, with steep discounting (i.e., delayed outcomes holding little value) considered a form of impulsive decision-making. Reward and punishment learning refers to how people adjust their decision behavior in response to rewards and punishments, with deficits indicating atypical reward and punishment processing related to clinical outcomes such as substance abuse. In this study, we tested the association between delay discounting and reward and punishment learning among adults (N = 299) from the United States that completed an intertemporal choice task and the play-or-pass IGT. We fit a joint model to the task data to simultaneously estimate parameters from the hyperbolic discounting model (for the intertemporal choice task) with parameters from a reinforcement learning model (for the play-or-pass IGT). The joint model also included parameters capturing cross-task associations between model parameters. Our results indicate that steep delay discounting is negatively associated with punishment learning. That is, people who do not value delayed outcomes also show less learning from punishment. Implications of these findings are discussed in the context of improving our assessment of reward valuation and reward and punishment learning as constructs relevant to mental illness.
Dynamic social interactions and feedback are crucial for understanding others’ emotions, particularly when confronted with contradictory emotional cues. Alexithymia, a condition that co-occurs with many psychiatric disorders, is characterized by impairment in emotional processing. However, computational mechanisms by which it alters social inferences based on feedback cues remain unexplored. To examine this, 60 participants with low and high levels of alexithymia completed an emotional learning task involving contradictory social (verbal and visual) cues to infer targets’ emotions. Computational analyses, including bin-based, reinforcement learning, and drift-diffusion modeling, revealed how alexithymia alters latent parameters that govern value updating and choice. Individuals with high alexithymia demonstrated lower accuracy in learning from social feedback. Drift diffusion analysis revealed a perceptual bias toward the visual cue, higher drift rates in the visual-correct condition, and greater evidence accumulation to infer others' emotions. These findings suggest that individuals with high alexithymia exhibit impaired social learning and difficulty with decision-making in situations with conflicting social information, with computational modeling quantifying the latent processes involved and advancing mechanistic targets for computational psychiatry.
Processing uncertainty may be pathognomonic (characteristic of a disease) for some psychiatric conditions. Some people expect the world to change, even when it doesn’t. This tendency is central to paranoia, where individuals often anticipate threat or change without clear evidence. But what determines whether these beliefs translate into behavior? One possibility is that metacognitive structure – the coherence and depth with which one articulates their own thinking – acts as a buffer. An agent may endorse a belief but have sufficient accessory hypotheses to insulate it from action. To test this, we used metacognitive prompting in GPT-4 to score individual reflections on open-ended questions (e.g., did you use any particular strategy?) after completing a probabilistic reversal learning task. Individuals with higher paranoia demonstrate lower metacognitive structure (t = 5.98, p < 0.001), with metacognition attenuating the relationship between volatility belief and switching behavior (Δ = –15 pp, p < 0.001) even after controlling for reflection verbosity and general cognitive ability. These findings suggest that metacognition protects against uncertainty-driven instability, pointing to a key mechanism by which reflection protects against cognition under change. This work provides a novel framework to measure metacognition from behavioral task debrief questions.
Both obsessions and paranoia are characterized by cognitive inflexibility, particularly in uncertain environments. Yet, differential diagnosis is challenging and limited to clinical interviews and self-report symptom questionnaires. We predicted that obsessions and paranoia would be associated with distinct patterns of behavior in our well-established probabilistic reversal learning (PRL) task. Probabilistic reversal learning involves updating beliefs about rewards when contingencies change. Obsessions and paranoia have been linked to excessive switching behaviors during reversal learning, although some report perseveration in patients with OCD and those with schizophrenia. Here, we analyze data gathered from the general population to assess the associations between obsessions, paranoia, and PRL task performance. Using a novel computational method – Bayesian Gaussian Graphical Modeling combined with a Hierarchical Gaussian Filter – we distinguish the impacts of paranoia and obsession on reversal learning – despite their significant correlation. We find that win-switching arises in paranoia from deficits in learning about uncertainty in the global task structure, whereas excessive switching in OCD arises from challenges in learning about uncertainty at the local choice level.
Objective:Identifying obsessive-compulsive disorder (OCD) using brain data remains challenging. Resting-state electroencephalography (EEG) offers an affordable and noninvasive approach, but identifying predictive signals in EEG data has met with little success, even with the application of traditional machine learning methods. We explored whether convolutional neural networks (CNNs) applied to EEG time-frequency representations can distinguish individuals with OCD from healthy controls. Method:We collected resting-state EEG data from 20 unmedicated participants (10 with OCD, 10 healthy controls). Four-second EEG segments were transformed into time-frequency representations. We then trained a 2D CNN using a leave-one-subject-out cross-validation framework to perform subject-level classification and compared its performance to a more traditional support vector machine (SVM) approach. Next, using multimodal fusion, we examined whether adding clinical and demographic information improved classification. Results:The CNN classifier achieved high subject-level performance, distinguishing individuals with an accuracy of 85.0% and an area under the curve (AUC) of 0.88. This significantly outperformed the SVM baseline, which performed no better than chance (45.0% accuracy, AUC: 0.47). A subsequent multimodal analysis revealed that clinical and demographic variables did not contribute any additional independent information. Conclusion:CNNs applied to resting-state EEG show promise for identifying OCD, outperforming traditional machine learning methods. These findings highlight the potential of deep learning to uncover complex, diagnostically relevant patterns in neural data. While limited by sample size, this work supports further investigation into multimodal models for psychiatric classification, warranting replication in larger, more diverse samples.
Anxiety disorders are chronic, pervasive, and debilitating; characterised by a persistent or exaggerated response to distal or abstract threats. Impaired threat discrimination (distinguishing safe from threatening stimuli) and impaired threat extinction (learning a once threatening stimulus is now safe), are known risk factors in the development and persistence of anxiety disorders. These effects can be experimentally elicited through fear conditioning. First, repeated trials of paired aversive and neutral stimuli are delivered during a fear acquisition phase, followed by repeated trials with no aversive stimuli in a fear extinction phase. The effects are typically measured through comparison of end-phase data points, or simple descriptive or statistical models. Computational modelling, by contrast, can offer a hypothesis-driven, trial-by-trial mechanistic account of fear conditioning. This unmasks within subject task variance by estimating the rate of threat learning, safety learning, and threat extinction, examining individual differences in the cognitive mechanisms behind anxiety. A normative sample (n = 145) underwent a differential fear conditioning task on a bespoke smartphone app, in addition to completing an anxiety severity measure (GAD-7). Computational models fitted to task data estimated learning rates. Whilst the threat learning rate showed no association, the threat extinction and safety learning rates showed small negative associations with anxiety severity (ρ = –0.22, p = 0.01 & ρ = –0.21, p = 0.01 respectively). These findings are in keeping with prior studies using traditional analytical approaches, and indicate that anxious individuals are not quicker to develop fear of a stimulus, but take more time than their non-anxious counterparts to learn that a stimulus is safe. This study strengthens the evidence for impairments in fear extinction in those with anxiety, and the importance of learning rates as an index of anxiety severity, a previously hidden cognitive mechanism underlying anxiety persistence.
Depression is a prevalent psychiatric condition that commonly emerges in adolescence and young adulthood and is associated with reward processing abnormalities. The Probabilistic Reward Task (PRT) is widely used to investigate the impact of depression on reward processing, but prior studies have not comprehensively addressed the reinforcement learning and decision-making mechanisms involved in the task. In 726 adolescents and young adults with varying levels of depression, we collected PRT data and applied a novel computational model with response-outcome learning and evidence accumulation processes to provide new insights into the cognitive processes implicated in depression. Compared to participants with no history of psychopathology, those with depressive disorders showed reduced impact of learned response values on decision bias toward the more frequently rewarded action. In addition, higher levels of anhedonia were associated with slower evidence accumulation during decision-making. Together, these findings improved our understanding of the reinforcement learning and decision-making mechanisms assessed by the PRT and their associations with depression.
When we smile, we expect that others will smile back. When one's smile is not reciprocated, these expectations are violated, producing prediction error signals in the brain. Prediction error signals may be experienced as aversive, disincentivizing smiling. Social smiling is impaired in psychotic disorders suggesting increased sensitivity to unreciprocated smiles. We developed the Incongruent Facial Emotion task to probe responses to unreciprocated smiles. Healthy controls and persons with schizophrenia or schizoaffective disorder voluntarily smiled, after which they viewed a stimulus face with a happy or angry expression. Brain activations were quantified with functional magnetic resonance imaging. Greater illness severity was associated with reduced smile amplitude. Across both groups, viewing an incongruent stimulus after initiating a smile activated the bilateral anterior insulae and right supplementary motor cortex. Brain activations in the left middle occipital and left superior frontal gyri were greater in the clinical group. The anterior insula response to incongruent facial reactions was significantly greater in more severely ill clinical participants. Dynamic causal modelling suggests that incongruent stimuli reduce tonic self-inhibition in the anterior insula, and that this disinhibition is enhanced by illness severity. The results suggest that the anterior insula processes affective prediction errors and sends feedback to supplementary motor areas to alter behavioural responses. The underlying brain circuits are enhanced in clinical participants with severe illness, suggesting new avenues to understand affective blunting in psychotic disorders.
Background: The Pavlovian go/no-go task is commonly used to measure individual differences in Pavlovian biases and their interaction with instrumental learning. However, prior research has found suboptimal reliability for computational model-based performance measures for this task, limiting its usefulness in individual-differences research. These studies did not make use of several strategies previously shown to enhance task-measure reliability (e.g., task gamification, hierarchical Bayesian modeling for model estimation). Here we investigated if such approaches could improve the task’s reliability. Methods: Across two experiments, we recruited two independent samples of adult participants (N=103, N=110) to complete a novel, gamified version of the Pavlovian go/no-go task multiple times over several weeks. We used hierarchical Bayesian modeling to derive reinforcement learning model-based indices of participants' task performance, and additionally to estimate the reliability of these measures. Results: In Experiment 1, we observed considerable and unexpected practice effects, with most participants reaching near-ceiling levels of performance with repeat testing. Consequently, the test-retest reliability of some model parameters was unacceptable (range: 0.379–0.973). In Experiment 2, participants completed a modified version of the task designed to lessen these practice effects. We observed greatly reduced practice effects and improved estimates of the test-retest reliability (range: 0.696–0.989). Conclusion: The results demonstrate that model-based measures of performance on the Pavlovian go/no-go task can reach levels of reliability sufficient for use in individual- differences research. However, additional investigation is necessary to validate the modified version of the task in other populations and settings.
Intolerance of Uncertainty (IU) is a transdiagnostic factor in psychological disorders, yet its underlying psychological mechanisms remain unclear. To close this gap, we first identify three potential mechanisms from existing definitions of IU: (1) negativity overweighting, (2) probability distortion, and (3) information deficit aversion. Second, we demonstrate how these mechanisms map onto well-established preference patterns in decision making under uncertainty as captured by Cumulative Prospect Theory: (1) loss aversion, (2) nonlinear probability weighting, and (3) the description–experience (DE) gap. Third, we conduct an affective decision-making experiment to investigate the relationship between self-reported IU and these preference patterns, as measured with individually estimated parameters of cumulative prospect theory. In the study, 100 participants made 120 choices between hypothetical painkillers with different probabilistic side effects. Half of the choices were made in a description condition, where all information was provided upfront; the other half in an experience condition, where participants acquired information through sampling. Trait IU was measured with a questionnaire. Participants overweighed side effects relative to treatment benefits (loss aversion), overestimated the probability of unlikely negative outcomes (increased nonlinear probability weighting), and their probability weighting patterns differed between the experimental conditions (DE gap). However, their preference patterns did not correlate with IU scores. Possible explanations are that the task did not effectively establish an affective context with real consequences for behavior, or that disorder-specific processes were not captured in our community sample. These findings highlight the need for a precise definition of IU and suggest avenues for designing tasks that enable a better understanding of IU.
Background:Parental capacity to learn from infant responses is a fundamental component of early dyadic interactions. However, the precise cognitive processes involved in these interactions and how these processes are influenced by mental health difficulties remain unclear. Methods:We investigated the computational basis of learning and decision-making in males and nulliparous females (Study 1) and pregnant participants enrolled in a cohort study (Study 2), using a two-armed bandit task adapted to simulate playful interactions with an infant. Participants chose between two competing bandits (i.e., two toys) with different underlying nominal probabilities for three outcomes (i.e., infant sad, neutral, and happy facial expressions). In Study 1, we manipulated the baseline emotional context of the task (i.e., the infant started either happy or sad) to investigate its effect on the processing of emotional feedback and decision-making. In both studies, we explored whether individual differences in mental health and personalities difficulties associated with variation in parameters. Results:In Study 1, the emotional context manipulation influenced both learning rates and how neutral outcomes were evaluated. Participants starting with a happy infant exhibited faster learning and a more negative evaluation of neutral outcomes compared to those starting with a sad infant. In Study 2, participants reporting higher levels of personality difficulties and antenatal depressive symptoms showed reduced learning rates. These associations were weaker in Study 1. Conclusions:Our findings provide novel evidence regarding the role of the emotional context in learning and decision-making processes. For parents with depressive symptoms and personality difficulties, dampened responsivity to emotional feedback and inflexibility in updating beliefs about the values of actions may underlie fewer sensitive behaviours when interacting with their infants.
Anorexia nervosa (AN) is a severe eating disorder, marked by persistent changes in behaviour, cognition and neural activity that result in insufficient body weight. Recently, there has been a growing interest in using computational approaches to understand the cognitive mechanisms that underlie AN symptoms, such as persistent weight loss behaviours, rigid rules around food and preoccupation with body size. Our aim was to systematically review progress in this emerging field. Based on articles selected using systematic and reproducible criteria, we identified five current themes in the computational study of AN: 1) reinforcement learning; 2) value-based decision-making; 3) goal-directed and habitual control over behaviour; 4) cognitive flexibility; and 5) theory-based accounts. In addition to describing and appraising the insights from each of these areas, we highlight methodological considerations for the field and outline promising future directions to establish the clinical relevance of (neuro)computational changes in AN.
Developing precise, innocuous markers of psychopathology and the processes that foster effective treatment would greatly advance the field's ability to detect and intervene on psychopathology. However, a central challenge in this area is that both assessment and treatment are conducted primarily in natural language, a medium that makes quantitative measurement difficult. Although recent advances have been made, much existing research in this area has been limited by reliance on previous-generation psycholinguistic tools. Here we build on previous work that identified a linguistic measure of "psychological distancing" (that is, viewing a negative situation as separated from oneself) in client language, which was associated with improved emotion regulation in laboratory settings and treatment progress in real-world therapeutic transcripts (Nook et al., 2017, 2022). However, this formulation was based on context-insensitive word count-based measures of distancing (pronoun person and verb tense), which limits the ability to detect more abstract expressions of psychological distance, such as counterfactual or conditional statements. This approach also leaves open many questions about how therapists' - likely subtler - language can effectively guide clients toward increased psychological distance. We address these gaps by introducing the use of appropriately prompted large language models (LLMs) to measure linguistic distance, and we compare these results to those obtained using traditional word-counting techniques. Our results show that LLMs offer a more nuanced and context-sensitive approach to assessing language, significantly enhancing our ability to model the relations between linguistic distance and symptoms. Moreover, this approach enables us to expand the scope of analysis beyond client language to shed insight into how therapists' language relates to client outcomes. Specifically, the LLM was able to detect ways in which a therapist's language encouraged a client to adopt distanced perspectives-rather than simply detecting the therapist themselves being distanced. This measure also reliably tracked the severity of patient symptoms, highlighting the potential of LLM-powered linguistic analysis to deepen our understanding of therapeutic processes.
Psychiatric disorders are highly heterogeneous and often co-morbid, posing specific challenges for effective treatment. Recently, computational modeling has emerged as a promising approach for characterizing sources of this heterogeneity, which could potentially aid in clinical differentiation. In this study, we tested whether computational mechanisms of decision-making under approach-avoidance conflict (AAC) – where behavior is expected to have both positive and negative outcomes – may have utility in this regard. We first carried out a set of pre-registered modeling analyses in a sample of 480 individuals who completed an established AAC task. These analyses aimed to replicate cross-sectional and longitudinal results from a prior dataset (N = 478) – suggesting that mechanisms of decision uncertainty (DU) and emotion conflict (EC) differentiate individuals with depression, anxiety, substance use disorders, and healthy comparisons. We then combined the prior and current datasets and employed a stacked machine learning approach to assess whether these computational measures could successfully perform out-of-sample classification between diagnostic groups. This revealed above-chance differentiation between affective and substance use disorders (balanced accuracy > 0.688), both in the presence and absence of co-morbidities. These results demonstrate the predictive utility of computational measures in characterizing distinct mechanisms of psychopathology and may point to novel treatment targets.
Eating disorders (EDs) are characterised by intense concerns about food and weight. These concerns are linked to changes in decision-making, such as persisting with actions that are no longer rewarding. For example, individuals might engage in long exercise sessions or time-consuming body checking practices, despite limited benefits. This study tested whether people with subclinical ED symptoms show increased persistence due to altered decision-making processes. Specifically, we postulated a shift in internal thresholds for making different decisions in EDs, which change the balance between exploitation and exploration. A subclinical group with heightened concerns about eating (sED; N = 44) and a healthy control group (HC; N = 56) completed a foraging task, in which an option on screen was exploited for reward. With each decision to exploit, reward feedback decreased and participants had to decide when to move on to a new option. Each block was time limited to 7.5 minutes. Behavioural persistence was measured as the number of seconds spent exploiting each option. Decision thresholds were measured when deciding to move on, as the counterfactual reward that would have been received for an exploit action. We predicted that the sED group would show increased persistence and decreased decision thresholds (i.e. lower counterfactual reward when deciding to move on) in comparison to the HC group. We found no evidence for these predictions. Instead, exploratory analyses showed that the sED group exhibited progressively faster response times (RTs) when approaching the time limit for each block. This increase in motor vigour was correlated with the severity of eating disorder symptoms from a range of traditional diagnostic categories. Our results point to changing motor vigour as a potential transdiagnostic marker of ED tendencies.
In value-based decision-making there is wide behavioural variability in how individuals respond to uncertainty. Maladaptive responses to uncertainty have been linked to a vulnerability to mental illness, for example, between risk aversion and affective disorders. Here, we examine individual differences in risk sensitivity when subjects confront options drawn from different value distributions, where these embody the same or different means and variances. In simulations, we show that a model that learns a distribution using Bayes' rule and reads out different parts of the distribution under the influence of a risk-sensitive parameter (Conditional Value at Risk, CVaR) predicts how likely an agent is to prefer a broader over a narrow distribution (pro-variance bias/risk-seeking) under the same overall means. Using empirical data, we show that CVaR estimates correlate with participants' pro-variance biases better than a range of alternative parameters derived from other models. Importantly, across two independent samples, CVaR estimates and participants' pro-variance bias negatively correlated with trait rumination, a common trait in depression and anxiety. We conclude that a Bayesian-CVaR model captures individual differences in sensitivity to variance in value distributions and task-independent trait dispositions linked to affective disorders.
Computational (generative) modelling of behaviour has considerable potential for clinical applications. In order to unlock the potential of generative models, reliable statistical inference is crucial. For this, Bayesian workflow has been suggested which, however, has rarely been applied in Translational Neuromodeling and Computational Psychiatry (TN/CP) so far. Here, we present a worked example of Bayesian workflow in the context of a typical application scenario for TN/CP. This application example uses Hierarchical Gaussian Filter (HGF) models, a family of computational models for hierarchical Bayesian belief updating. When equipped with a suitable response model, HGF models can be fit to behavioural data from cognitive tasks; these data frequently consist of binary responses and are typically univariate. This poses challenges for statistical inference due to the limited information contained in such data. We present a novel set of response models that allow for simultaneous inference from multivariate (here: two) behavioural data types. Using both simulations and empirical data from a speed-incentivised associative reward learning (SPIRL) task, we show that models harnessing information from two different data streams (binary responses and continuous response times) ensure robust inference (specifically, identifiability of parameters and models). Moreover, we find a linear relationship between log-transformed response times in the SPIRL task and participants' uncertainty about the outcome. Our analysis illustrates the benefits of Bayesian workflow for a typical use case in TN/CP. We argue that adopting Bayesian workflow for generative modelling helps increase the transparency and robustness of results, which in turn is of fundamental importance for the long-term success of TN/CP.
