Addressing Altered Anticipation as a Transdiagnostic Target through Computational Psychiatry

Abstract

Anticipation of future experiences is a crucial cognitive function impacted in various psychiatric conditions. Despite significant research advancements, the mechanisms underlying altered anticipation remain poorly understood, and effective targeted treatments are largely lacking. This review proposes an integrated computational psychiatry approach to address these challenges. We begin by outlining how altered anticipation presents across different psychiatric conditions, including schizophrenia, major depressive disorder, anxiety disorders, substance use disorders and eating disorders, and summarizing the insights gained from extensive research using self-report scales and task-based neuroimaging, despite notable limitations. We then explore how emerging computational modeling approaches, such as reinforcement learning and anticipatory utility theory, could overcome these limitations and offer deeper insights into underlying mechanisms and individual variations. We propose that integrating these interdisciplinary methodologies can offer comprehensive transdiagnostic insights, aiding the discovery of new therapeutic targets and advancing precision psychiatry.

Introduction

People spend a long time anticipating future experiences, often longer than the duration of the experiences themselves. This anticipation elicits a range of emotions that impact subjective well-being and decision-making(1–3). For instance, merely thinking about a planned trip to a beach resort can make people happy, leading them to schedule the trip weeks ahead to have more time to enjoy the anticipation(1,2). Conversely, the anticipation of an unavoidable painful dental procedure can be dreadful, prompting many people to want to complete it as soon as possible (2,4–6).

Alterations in anticipation manifest differentially across psychiatric conditions (Figure 1). For instance, individuals experiencing anhedonia (diminished pleasure) in the context of schizophrenia or major depressive disorder (MDD) often report difficulty enjoying the anticipation of future rewards (anticipatory anhedonia), though they may still enjoy the actual rewarding experience itself(7,8). Other conditions, such as anxiety disorders, involve unbearable emotions when anticipating future experiences(9). Individuals with substance use disorders (SUD) may intensely anticipate the act of using substances or dread withdrawal, which can play a significant role in their condition(10). In eating disorders (ED), such as anorexia nervosa (AN), individuals might positively anticipate weight loss but experience intense dread at the thought of eating or gaining weight(11).

Figure 1. Altered anticipation as a transdiagnostic target.

Altered anticipation manifests across a variety of psychiatric disorders, including in schizophrenia, major depressive disorders, anxiety disorders, substance use disorders, and eating disorders. Symptoms associated with anticipation, such as anticipatory anhedonia, anticipatory anxiety, craving, and excessive dread are often expressed transdiagnostically across the traditional diagnostic boundaries. Because of its clinical significance, anticipation has been one of the most extensively studied cognitive processes. Established neurocognitive approaches include self-report questionnaires, monetary incentive delay (MID) task and the neutral predictable unpredictable (NPU) threat test. In addition, computational models that quantitatively describe anticipatory processes could offer additional valuable insights. The new digital platforms, such as online recruitment frameworks and machine learning techniques, now allow us to integrate all these approaches and gain transdiagnostic insights. The new approach may help us identify meaningful individual variation and individualized transdiagnostic treatment targets for altered anticipation. Symptoms potentially associated with altered positive and negative anticipation are presented in green and purple, respectively. Black lines indicate symptoms typically associated with psychiatric disorders, while grey lines denote less common associations.

Significant advancements in neurocognitive research methods, such as self-report scales and task-based neuroimaging, have provided robust insights into the mechanisms underlying altered anticipation(12–17). However, these methods have notable limitations. Emerging computational modeling approaches, such as reinforcement learning theory and anticipatory utility in behavioral economics(1,2), show promising potential for addressing existing methods’ limitations, though these are not yet widely adopted in clinical research. Given the complexity of altered anticipation across psychiatric conditions, an effective path forward would integrate these interdisciplinary methodologies. The recent advancement of digital technologies, such as large-scale online study platforms and machine learning methods to analyze big data, can facilitate such integration. This computational psychiatry approach has the potential to offer comprehensive transdiagnostic insights, aiding in the discovery of new therapeutic targets and advancing precision psychiatry (Figure 1).

This paper is structured as follows: First, we provide an overview of how altered anticipation manifests across various psychiatric conditions. Given their complexity, our discussion focuses on specific aspects relevant to altered anticipation, rather than providing an exhaustive review of the conditions. Next, we explore emerging computational modeling approaches and discuss how their integration with other established neurocognitive methods could enhance our understanding of altered anticipation.

Altered anticipation across psychiatric conditions

i. Anticipatory anhedonia as diminished positive anticipation.

Anhedonia, a loss of pleasure, is a disabling symptom across neuropsychiatric conditions, including schizophrenia, MDD, bipolar spectrum disorders, and Parkinson’s disease (7,8,13). There are at least two known subtypes of anhedonia: anticipatory and consummatory anhedonia(7). Individuals with anticipatory anhedonia do not experience pleasure from anticipating future rewards, while individuals with consummatory anhedonia do not experience pleasure from experiencing the reward itself. While both forms appear in MDD and bipolar disorder (18,19), schizophrenia is more commonly linked to anticipatory anhedonia (8) (however, see (20), where the authors discuss the challenge of using self-report questionnaires to examine consummatory anhedonia).

One outstanding question is why some individuals display anticipatory anhedonia while their consummatory pleasure remains mostly intact. A classic account attributes this to “wanting” (for anticipatory) versus “liking” (for consummatory) circuits (7,8,12,18). The wanting circuit centrally involves mesolimbic dopamine, while the liking circuit is traditionally associated with dopamine-independent “hedonic hotspots”(12). Neuroimaging supports mesolimbic involvement in schizophrenia (21), but dopamine’s role in anticipatory anhedonia remains unclear. While dopamine-targeting antipsychotics treat positive symptoms (e.g., hallucinations), they fail to improve negative symptoms, including anhedonia(22,23).

Recent studies suggest anhedonia (note: not necessarily anticipatory anhedonia) is central to schizophrenia’s symptom network (24), particularly its association with reduced cognitive function, which are debilitating and largely untreatable (25). There is a robust clinical correlation between anhedonia and cognitive symptoms, such as episodic memory deficits in schizophrenia(8,24,25) and MDD(26,27), though the mechanism underlying this co-occurrence remains unclear.

One key brain structure involved may be the hippocampus. Hippocampal volume reduction is implicated in MDD (28) and is associated with memory deficits (29) and negative symptoms(30) in schizophrenia. Altered NMDA receptors in the hippocampus have also been linked to negative symptoms and cognitive deficits(31,32), and hippocampal alterations have been proposed to drive dysfunctional midbrain dopamine release(33). Individuals with schizophrenia show stronger temporal discounting(34) , impaired prospection(35), novelty responses(36,37), and mental imagery (38,39), all of which involve the hippocampus. As we will discuss, recent computational modeling also supports the role of the hippocampus in anticipation(40). These findings suggest that anticipatory anhedonia and cognitive deficits could result from common neural alterations involving the hippocampus, but further studies are required to understand the roles of subregions(41).

ii. Anxiety and dread as negative anticipation impacting information seeking and avoidance.

While diminished positive anticipation can be associated with anhedonia, excessive negative anticipation may be associated with anxiety and dread(1,9,42). These sustained emotions are often associated with reactions to unpredictable stimuli ((43), although see (44) for a more nuanced view) and distinguished from fear - a fast, transient response. Excessive dread is a debilitating feature of anxiety disorders and obsessive-compulsive disorder (OCD), and often co-occurs with affective, personality, and psychotic disorders(45,46).

A key behavioral response to anxiety is avoidance. This includes not only avoiding harm (47,48) but also avoiding information about potential harm(49). The latter, though relatively underexplored, may be critical in understanding individual differences in coping with anxiety or negative anticipatory feelings. For instance, some individuals may choose not to receive information about a potential cancer diagnosis(50,51). Information may amplify fear and anxiety(52,53) or alleviate anxiety before medical procedures(54). Having information about future harm can sharpen attention and intensify dread(55), while a lack of information can cause uncertainty and anxiety(1,56). If a person experiences a sense of control with information, e.g., if they believe that they can reduce the pain through preparation(57), advance information about future harm could reduce dread. Relatedly, anxiety prior to aversive intervention can enhance pain response to that intervention(58,59). Consequently, those who would feel less anxious with more (or less) knowledge may actively seek (or avoid) information(54).

Information seeking behavior in psychiatric conditions is underexplored. One study suggests that people with social anxiety seek less information about social ratings than neurotypical controls(60), while another suggests people with OCD gather information more efficiently (61). The amount of uncertainty(62), anxiety(63), and information content may impact avoidance behavior. Because individuals with anxiety and depression often report learned helplessness and a sense of certainty regarding adverse outcomes(64,65), studying how a sense of control may mediate information avoidance could offer valuable insights.

Anxiety impacts multiple brain regions, with the ventral hippocampus (rodents) and anterior hippocampus (humans) linked to evoked anxiety, while the posterior hippocampus relates to trait anxiety (67,68). Hippocampal alterations may underlie anxiety disorders and cognitive deficits(69). Anxiety and loss anticipation also involve reduced ventromedial prefrontal and ventral striatum activity, with increased insula and anterior cingulate activity(5,16,42,70,71).

iii. Potential interplays between positive and negative anticipation in substance use and eating disorders.

In the previous two subsections we have discussed the implications of alterations in positive (anhedonia) and negative (dread, anxiety) anticipation, respectively. Here we discuss alterations involving both positive and negative valence in the context of substance use disorder (SUD) and eating disorders.

Substance use disorders (SUD):

SUD is a devastating mental illness with one of the highest mortality rates among all psychiatric conditions (72). Early SUD stages can involve heightened pleasure from substance-use(73), accompanied by enhanced dopamine release in response to substance-related cues (incentive-sensitization)(12) and consumption(74,75). Over time, substance tolerance(76) can attenuate the hedonic experience, emphasizing negative experiences from abstinence(77). Neurotransmitter imbalance(78) and anxiety-like states (79) may promote hyperkatifeia (80) and craving (77). These suggest complex interplays of positive and negative anticipatory feelings. Early stages can involve enhanced positive anticipation of substance use(81,82), while later stages emphasize the dread of being unable to use substances and the prospects of worsening withdrawal symptoms (10).

The factors influencing SUD are complex, potentially involving top-down and bottom-up processes(83). Impulsive traits, increased delay discounting(84,85) and altered striatal D2 receptors are associated with addictive behavior(83,86–89) . Individuals with SUD show increased neural responses to cues associated with substances (90,91) and report enhanced anticipatory pleasure (81,82), especially when use is guaranteed(92). Repeated use has a pharmacological impact, exemplified by conditioned tolerance, where metabolic adaptation prepares the body for substance intake (76). Context also matters. Drug use in unfamiliar settings increases overdose risk, as seen in both humans and other animals (93–95).

Similarly to anhedonia and anxiety, the hippocampus may be crucial(10,96). Enhanced hippocampal activity can foster drug-context associations (73), leading to heightened cue-induced craving(97). Enhanced cognitive function in early nicotine users is well-documented(98), but withdrawal may impair hippocampus-dependent cognition (73,96,99).

Eating disorders (ED):

ED involves cognitive, emotional, and behavioral difficulties regarding food, body image, and eating. Anorexia nervosa (AN) is characterized by severe self-starvation and a high mortality rate (100,101). Bulimia nervosa (BN) involves recurrent binge eating and compensatory behaviors (purging), whereas binge-eating disorder (BED) is characterized by binge eating without purging(102).

Standard mechanistic accounts for ED often center around goal-directed control and habit(103). AN’s restrictive eating may arise from a strong goal-directed desire to lose weight (104), overriding hunger signals (11). Over time, restrictive behavior can become habitual (105,106) with decreased self-control(107). However, this account sometimes struggles to capture the complexity of ED. BED and BN are associated with both decreased self-control and increased habitual behaviors around binging(108,109), yet treatments that focus on unrestricted eating are effective (110). Further, patients sometimes transition between restrictive-type AN and binge eating behavior(111) (though see (101)).

This transition between different stages of ED is a crucial, yet understudied feature that may be clarified by studying anticipation. The initial phase of restrictive eating in AN may be motivated by positive anticipation of weight loss(112,113), consistent with reduced temporal discounting(114) and attenuated impulsivity(115). Like in SUD, food restriction can alter dopamine function(116), increasing transporter levels(117) and generating a sensitized system. Over time, this positive anticipation of future weight loss may transition to (or co-occur with) negative anticipation (dread) of gaining weight and its social ramifications(11,118), accompanying a strong avoidance of high-calorie food. Individuals with AN have shown enhanced (aversive) anticipatory response to high-calorie food stimuli(119). Similarly to schizophrenia, individuals with AN and BN report reduced anticipatory but intact consummatory pleasure(115).

These interplays between positive and negative anticipation may arguably be oversimplified descriptions. However, this viewpoint may offer potentially shared neural mechanisms underlying SUD and ED, given the similarity and comorbidities (120,121), including excessive dopamine release (122), and reduced hippocampal volumes(123,124). Further studies may illuminate shared mechanisms.

Approaches for understanding altered anticipation

Neurocognitive approaches

Advancements in neurocognitive research methodologies have provided robust transdiagnostic insights into the mechanisms underlying altered anticipation (Table 1). Self-report questionnaires are one of the most extensively used strategies in clinical psychology, with some designed to probe anticipatory traits (Figure 2A). Behavioral tasks, such as the Monetary Incentive Delay (MID)(15) or the Neutral Predictable Uncertain (NPU) Threat(125) tasks, have been further developed to explore at physiological and neural levels the anticipatory responses (Figure 2BC). However, these approaches struggle to capture more complex behaviors, e.g., the impact of anticipation on decision-making and the underlying computational mechanisms in anticipatory processing.

Table 1: Neurocognitive approaches to study anticipation.

Methodology	Description	Pros/ Cons	Examples of Clinical Implementation
Self-report Questionnaires	Self-report scales typically include items that inquire about an individual’s expectations, excitement, and motivation in response to potential rewards (Figure 2A). Classic metrics are mostly focused on consummatory pleasure, while second generation scales are designed to target anticipatory and consummatory pleasure separately (e.g. Temporal Experience of Pleasure, TEPS). Anticipation of negative experiences is in general less explored in self-report scales, with some exceptions focusing on anxiety by generated by future aversive events. Other relevant instruments have focused on attitudes towards uncertainty	Self-reports questionnaires are easy administer and can offer insights about individual anticipatory traits across psychiatric disorders, such as schizophrenia and eating disorders. Inconsistent findings and difficulty to delineate anticipatory vs. consummatory pleasure solely from questionnaires, since reporting consummatory pleasure demands various additional cognitive functions, such as imagination and memory.	Temporal Experience of Pleasure Scale (TEPS) in schizophrenia has indicated specific deficits to anticipatory pleasure processing. Intolerance to Uncertainty questionnaires administered in various psychiatric conditions, including generalized anxiety disorder OCD, and depression.
Behavioral tasks
Monetary Incentive Delay Task (MID)	Neuroimaging based paradigm, where participants receive a cue signaling that they will receive a monetary reward if they successfully press a button in response to a second cue displayed after a few seconds (Figure 2B). Functional magnetic resonance imaging (fMRI) response to the first reward-predictive cue is often considered a neural correlate of reward anticipation, for which regions in the mesolimbic dopamine circuit, including the ventral striatum and ventral tegmental area, as well as anterior cingulate are reliably implicated.	One of the most robust and influential approaches to measure anticipatory signals at a neural level. MID paradigm uses as the primary behavioral measure participant’s reaction time, which is unclear how it is associated with altered anticipation. Attention and sensory-motor processing could also affect the reaction time. The typical anticipatory period of the MID paradigm (a few seconds) makes it challenging to dissociate phasic prediction error from slow reward anticipation. Such a delineation may require a task with longer anticipatory periods and computational models that make quantitative predictions about anticipatory processes.	Individuals with schizophrenia show blunted responses in the ventral striatum compared to healthy volunteers. Individuals with major depressive disorders show larger anterior cingulate cortex response and similar or smaller ventral striatal response than healthy volunteers but the alteration can be normalized by selective serotonin reuptake inhibitor (SSRI) treatment.
Neutral Predictable Uncertain (NPU) Threat	Developed based on classical conditioning, the NPU threat test examines potentially distinct responses to predictable and unpredictable threats (Figure 2C). In this paradigm, fear is associated with a fast and transient response to discrete and certain aversive stimuli, whereas anxiety is associated with sustained emotional response to unpredictable stimuli. Fear and anxiety are assessed using the startle reflex, a measure of aversive state elicited by the unpleasant stimuli, e.g. electric shocks.	The startle is a cross-species metric extendedly used in experiments that can be implemented multiple populations, facilitating translational research. A limitation is the gap between the simple startle responses, or eye blinks, and complex cognitive behaviors manifested in psychiatric conditions, such as decision making.	Individuals with panic disorder and post-traumatic stress disorder present greater startle responses to unpredictable, but not predictable, aversive experiences relative to healthy volunteers. Inhibited startle response to unpredictable events has been associated with avoidance strategies and intolerance of uncertainty.

Figure 2. Classic approaches to study (altered) anticipation.

(A). Self-report questionnaires. There are a variety of self-report scales to examine altered anticipation. A question from the Temporal Experience of Pleasure Scale (TEPS)¹⁴ is shown as an example. (B). Monetary Incentive Delay (MID) task. The MID task is one of the most utilized neuroimaging tasks to examine anticipatory responses in the brain. Participants are presented with a cue indicating the magnitude of upcoming rewards. Participants need to press a button in response to a second cue that papers after a brief anticipation phase. fMRI response to the first cue has been used as a measure for neural correlates of anticipation, where the brain regions, including mesolimbic regions, are reliably activated. Reproduced from (16). (C). Neutral Predictable and Unpredictable Threat (NPU) test. The NPU test is an established lab-based task to examine anticipatory responses to potential negative outcomes, such as electric shocks or screaming voices. In the trials with the N condition, a cue signaling no upcoming aversive stimulus is presented to participants. In the P condition, a cue signaling an upcoming aversive stimulus is presented. In the U condition, a cue signaling a potential aversive stimulus is presented. Psychophysiological responses, most notably startle responses measured by electromyography (EMG) activity in the orbicularis oculi muscle, have been used to characterize anticipatory responses to these cues. Reproduced from(125).

Computational modeling approaches

We suggest that the limitations of the established approaches could be compensated by emerging computational modeling and behavioral paradigms. Here we describe two computational models: classic reinforcement learning models, which posit that hedonic value originates from the consumption of rewards, and the anticipatory utility model, which suggests that additional value arises from the anticipation of rewards (Figure 3).

Figure 3. Computational models of anticipation.

(A). A rewarding event (e.g., a holiday trip) can consist of a long anticipation and the actual experience. (B). In standard decision-making models, such as the reinforcement learning (RL) model, the total value of the reward experience is computed solely by the reward itself, indicated by the blue bar. When gaining advance information that increases the likelihood of the upcoming reward (e.g., winning plane tickets for a holiday), the model generates a prediction error (PE; the dotted orange bar), a key signal for reinforcement learning. (C). Midbrain dopaminergic (DA) neurons have been shown to capture this prediction error signal. (Left panel) In a conditioning task, DA neurons initially respond to the reward itself (R). (Right panel) After many trials, the DA neurons respond to the cue predicting upcoming reward (conditioned stimulus: CS). The DA neurons suppress their activities when the expected reward is omitted (no R). Reproduced from (129) (D). This prediction error signal has also been implicated in the brain regions, including the ventral striatum, in human fMRI experiments. This is consistent with findings from MID tasks. Reproduced from (130). (E). The Anticipatory Utility (AU) model. In contrast to standard models, the AU model proposes that the anticipatory process can generate its own value (e.g., the pleasure of imagining future holiday activities, indicated by the green area) in addition to reward consumption. In the model, prediction errors experienced at advance information can boost anticipation (indicated by the orange upward arrows). (F). The AU task. In each trial, participants choose between an immediate-information (“Find out now”) and a no-information target (“Keep it secret”). Before making the choices, participants see a pie chart and an hourglass signal the probability of a reward and the duration of a delay period until the reward or no-reward delivery, respectively. Once a target is selected, a symbolic image cue appears during the delay, followed by a reward or no reward (in this example, the reward is a video of cute animals). The immediate information target is followed by cues that predict the upcoming reward or no reward. The no-information target is followed by a cue with no information. The probability of reward and the duration of delays are the same regardless of the participant’s choice. (G). The AU model predicts increased preferences for immediate information at longer delay conditions (blue), which are validated experimentally (black)(40), while standard models (such as RL) predict no preference. (H). The neural circuitry underlying AU computation. Model-based fMRI analyses have suggested three distinctive regions orchestrate the AU computation. The functional coupling between the hippocampus and the dopaminergic midbrain has been suggested to boost the AU value encoded in the vmPFC(40). (I). The temporal dynamics of the fMRI signal in the vmPFC matched the model’s AU value signal during the delay period(40).

i. Reinforcement learning model

Reinforcement learning (RL) is a widely used framework in machine learning and neuroscience to describe how agents learn through trial-and-error to maximize future reward (or minimize punishment) (126,127). This model generally computes the expected value of various states or stimuli using reward prediction errors (RPE)(126,128), which signal the disparity between predicted and actual outcomes (Figure 3B). A positive RPE occurs when a reward is unexpected, while a negative RPE occurs when an expected reward is omitted. RPEs also arise from cues predicting future rewards or punishments (temporal difference errors)(126,128). This theoretical RPE signal has been shown to be consistent with the phasic responses in midbrain dopamine neurons(129) (Figure 3C) that project to the ventral striatum(127,130) (Figure 3D), supporting the model’s biological relevance.

RL models help explain computational variations in clinical populations (Table 2), including early SUD stages as the increased expected value of substances and associated cues(75,131), and ED’s altered learning patterns(132,133).

Table 2: Clinical implications of computational frameworks.

Symptom or condition	RL model implications	AU model implications
Anticipatory anhedonia	• Impaired reward sensitivity and choice variability among patients with anhedonia may impact anticipatory processes. • Blunted responses to feedback information in individuals with anhedonia can be reflected by reduced learning rates. • Lower reward values (e.g., generated by lower learning rate for +RPE relative to −RPE, biased sampling of negative over positive experiences when constructing value) could also lead to altered reward anticipation.	• Reduced value of the anticipatory utility could generate weaker preference for advanced information (Fig. 4A):  ○ Lack of hippocampal boosting generating a weaker preference for advance information.  ○ Smaller dopaminergic response to anticipatory cues (consistent with MID tasks).  ○ Attenuated AU signal in the vmPFC. • Model-based fMRI analyses of the AU paradigm could identify distinct neural alterations underlying anhedonia across individuals, i.e., hippocampus alone, versus altered coupling between dopaminergic midbrain and the hippocampus. • Potential circuitry alteration involving the hippocampus could explain the known association between anticipatory anhedonia and cognitive deficits in the context of schizophrenia, as the hippocampus plays a major role in cognitive functions, such as episodic memory.
Negative anticipation and anxiety	• Biased anticipation of aversive events by enhanced punishment-related learning or reduced reward-sensitivity. • Stressful contexts may increase Pavlovian avoidance bias in anxious individuals, reflected on RL model parameters.	• Study anxiety qualitatively using the AU task with negative outcomes, such as electric shocks or screaming voices, enables examining how people decide to cope with anxiety and dread (Fig. 4B).  ○ Those who experience stronger dread with information about future outcomes will avoid advance information.  ○ Those who are afraid of uncertainty may seek advance information in the aversive AU task. • These behaviors can be captured by different parameter regimes of the AU model. • Test the neural correlates of anxiety or dread signals (which remain largely unknown) using the long delays during the AU task.
Substance use disorders	• Over-expressed prediction error at substance intake (reward consumption) generates an inflated expected value about this stimulus, leading to addiction. • Repeated substance use boosts the expected value of substances and substance-associated cues, which triggers enhanced expectation of future consumption. • Substance availability expectations and abstinence can modulate the learning from positive and negative outcomes.	• AU task behavior may change over documented illness progression, helping to identify individual variation and neural basis across different SUD states (Fig. 4C).  ○ Early stage may show unbounded boosting by a strong dopaminergic prediction error at the moment of predictive cues (PE Boost) could lead to an over-expressed anticipatory utility. This could reflect bottom-up effects that influence the positive motivational state in the early stages of SUD (e.g. people that experience enhanced pleasure during the anticipation of upcoming certain substance use)  ○ In late stages the dread of being unable to use substances and the prospects of worsening withdrawal symptoms can have a large negative AU, driving more substance use. This negative anticipation may result from complex interactions of both cognitive (top-down) and metabolic (bottom-up) anticipations. • Strong reframing (tendency of people who perceived reward positively, to change perception of neutral outcome as negative) resulting in heterogeneous choice behavior. • Positive AU could arise from a bias towards identifying potential future reward (e.g., substance consumption), while the negative AU arises from potential future no-reward (e.g., unavailability of the substance). This would result in heterogeneous choice behavior in the AU task: those who experience a large positive anticipation prefer advance information, while those who experience a large negative anticipation avoid information.
Eating disorders	• People with AN may exhibit stronger goal-directed or model-based control. • Patients with AN may show decreased temporal discounting. • Conversely, people with AN may exhibit reduced model-based RL as symptoms progress. • People with BED and BN may have reduced model-based RL. • Value of high calorie stimuli is decreased in patients with restrictive eating disorders	• Characterize the complex anticipatory process that may take place in ED, including positive anticipation toward weight loss, motivating food avoidance and overexercise, and the negative anticipation by dread of regaining weight and associated social outcomes (Fig. 4C). • Individual variations in AU task behavior could delineate the competitive positive anticipation and negative anticipatory states, which may help to predict inflection points in eating disorders. • Information-seeking behavior could reflect treatment changes since some patients exhibit patterns of repeated checking of weight, while others intentionally avoid it. This behavior can change during treatment, during which many patients prefer having “blind weighing” to reduce their anxiety.

RL models can offer alternative interpretations of existing findings from MID tasks. Clinical populations often show blunted neural response to the reward predictive cues in the MID task(134–136), typically interpreted as a smaller reward anticipation. However, this could also indicate smaller RPEs(137,138), implying a larger reward expectation in the RL framework(139). Patients with MDD and schizophrenia have been reported to have a smaller response to estimated RPEs relative to other volunteers(140,141). Because RPEs and expected values are negatively correlated by definition(128,142), distinguishing these in simple tasks like the MID task is challenging. This highlights the need for an alternative task and a different computational modeling approach.

ii. Anticipatory utility model

The Anticipatory Utility (AU) theory is a computational framework that integrates anticipatory processes into value computation. Combining a computational model, behavioral task, and model-based neuroimaging, the AU framework enables the study of anticipatory processes and their effects on behavior, including potential alterations. Because this framework has not been fully explored in psychiatry, here we offer the comprehensive overview of the AU framework and its clinical implications (Figure 4 and Table 2).

Figure 4. Implications of AU framework: altered anticipation across psychiatric conditions.

(A). The AU model describes a range of positive anticipation processes, which are often referred to as savoring. A diminished positive AU value can capture anticipatory anhedonia, commonly expressed in the context of schizophrenia and major depression. (B). The AU model can describe negative AU value, called dread (e.g., waiting for a painful dental appointment). Advance information could reduce or enhance dread, reflected in the known individual variations in information seeking strategies. The people who can reduce their experience of dread with advance knowledge may seek information, while the people who experience enhanced dread with knowledge may avoid information. (C). While the anticipation of positive and negative experiences can coexist, the shifts in relative strengths between positive and negative anticipation may describe different stages of psychiatric conditions. Positive anticipation may be implicated in the early stages of substance use disorders (SUD) and anorexia nervosa (AN), where the future reward can be associated with, e.g., substance use and losing weight, respectively. Negative anticipation may be implicated in the progression of the conditions, e.g., the dread of being unable to use substances or withdrawal symptoms in SUD, and eating or gaining weight in AN. The actual relative dominance likely oscillates in a more complex manner.

The computational model:

the AU model is based on the mathematical model of the utility of anticipation developed in behavioral economics(2). To illustrate the idea, imagine being informed that you have won an opportunity to meet your favorite celebrity (e.g., Beyoncé). You are then asked: “When would you like the meeting to happen?”.

Standard RL models (138,143) predict a preference for an immediate meeting with the celebrity. This is because the model seeks to maximize the value of the meeting alone, which would decrease over time due to temporal discounting(144). However, studies have shown that many people would prefer to delay the meeting(2). To capture this, Loewenstein(2) formulated the AU theory, which incorporates the value arising from the anticipation of the future experience (Figure 3E). Formally, the total value in the AU model is the sum of the value of anticipation and the value of the reward(2):

$$TotalValue=\eta V^{anticipation}+V^{reward},$$

where the first term is the value arising from anticipation, η is the coefficient of anticipation, and the second term represents the value of the reward itself. This model reduces to a classic RL model at the limit of no anticipatory value (η=0)(40,55). The value of anticipation is computed as the total area under an exponentially increasing anticipation signal. Both anticipatory and reward consumption values follow standard temporal discounting, meaning that people may postpone a reward but not indefinitely(2). The value of anticipation arising from future reward at time T can be expressed as:

$$V^{anticipation}={\int }_0^T{e}^{-\gamma t}{Re}^{-\mu (T-t)}dt,$$

where μ is the ramping rate for anticipation, γ is the discounting rate, and R is reward value. This equation describes how anticipatory value dynamically evolves in time, thus making predictions about the exact time course of anticipation computed in the brain.

Inspired by observations in animal models(55,145), the AU model further suggests that people will experience a greater amount of AU after receiving information about future reward (e.g., news that a reward is arriving for sure) than with no such information (e.g., uncertain whether a future reward is arriving or not) (40,55). Such advanced information causes dopaminergic prediction errors, which in turn enhances the AU(55,145) (Figure 3E, PE Boost). Formally, this is expressed as:

$$\eta =c_0+c\times \left|\delta \right|,$$

where c₀ and c are constants which determine the anticipatory strength in each individual, and δ is the (anticipatory) prediction error, or the difference between the expected and actual value of cues computed from the AU model. The boosting constant c has been shown to correlate with the functional connectivity between the vmPFC and the hippocampus during anticipation(40). This augmentation is critical for capturing varieties of behavior that classic models struggle to account for, including non-instrumental information-seeking behavior, a core element of the AU experimental paradigm we will describe below (55,62,146).

The AU model can also cohesively explain behavior motivated by the anticipation of negative outcomes (dread(2)). For example, it explains why people would sometimes wish to have unavoidable distressing experiences (e.g., an electric shock) occur sooner rather than later in order to reduce the amount of dread experienced while waiting(4,5). Again, these behaviors cannot be captured by standard computational models, including classic RL models.

More generally, previous studies have shown that the anticipation of potential positive outcomes can involve a mixture of positive and negative anticipation(40,55) due to reframing:

$$V^{anticipation}={\eta }^{+}{V}^{Positiveanticipation}+{\eta }^{-}{V}^{Negativeanticipation}$$

where ${\eta }^{+}$ is the relative weight of positive anticipation (e.g., reward) and ${\eta }^{-}$ the weight of negative domain anticipation (e.g., no-reward).

Behavioral task:

The anticipatory utility task (AU task; Figure 3F) can test behavioral and neural predictions of the model in the context of information-seeking(40,55). The task, also referred to as the information-seeking task, is based on a series of experiments in animal models (145–147) and has been validated in human behavioral and neuroimaging experiments(40,55). In each trial, participants are informed about the chance of receiving an outcome after a certain time delay (e.g., a 50% chance of receiving a reward after 30 seconds). The participants are asked to choose from alternative options: “Find out now,” to immediately reveal if they will receive a reward or not, or “Keep it secret” to keep the outcome uncertain until they eventually receive it after a delay. Importantly, participants’ choices do not increase or reduce the chance of receiving a reward, i.e., they can only choose whether to gain advance information about the outcome or not. This choice also does not affect the delay, as participants still must wait out the full time-course of the trial. While this cue-reward association is similar to the one in the MID task, the AU task is distinct because it gives participants a choice, offering clear behavioral correlates. The task generally uses primary rewards (e.g., images) instead of monetary rewards, as anticipatory utility for monetary rewards has been found to be smaller(2). However, further research is needed.

Standard decision theories like RL models predict no preference in this task because the chance of receiving a reward is the same regardless of information choice. In contrast, the AU model predicts that participants should increase their preference for advance information about future outcomes under longer delay conditions. This occurs because, while the amount of AU is negligible when the delay is short, at longer delays people can experience a significantly greater AU value with advance information (through boosting with prediction error) than without advance information (no boosting). This key behavioral prediction has been validated across multiple studies(40,55) (Figure 3G). The findings are consistent regardless of whether subjects learned the experimental conditions (e.g., delays) through experience or explicit instructions(55).

Like other task-based frameworks, the AU framework has limitations. For example, participants may use heuristics to make their choices(148). This limitation can be mitigated by careful instructions and training sessions for participants to fully experience anticipation (149,150), as shown in previous studies on the AU task(40,55).

One alternative account for AU task behaviors is uncertainty reduction, where information carries positive value (146,151). However, this does not fully account for the observed valence- and delay-dependent information seeking. Participants choose less knowledge for undesirable or aversive outcomes (62) but prefer to receive more knowledge for longer outcome delays (55). The AU model can capture these behaviors, linking uncertainty reductions to prediction errors that boost anticipation (40,55). Another potential explanation is the modulation of discounting rate by prediction error or uncertainty(152), or altered time perception (55). Because this is a relatively new paradigm, further validation, including the test-retest reliability across various psychiatric conditions, is needed.

Neural circuits for AU value computation:

The AU model makes experimental predictions about the exact time-course of the anticipatory value signal and other signals arising from the model in the AU task, which can be tested in neuroimaging experiments using the model-based fMRI approach(40,153). Model-based fMRI analysis starts with fitting model parameters to participants’ choice behavior in the AU task. The model can generate predictions about the exact temporal dynamics of fMRI signals encoding the model’s key variables, including prediction errors from advance information and the dynamically evolving AU.

This model-based fMRI analysis applied to healthy volunteers performing the AU task has revealed a neural circuit consisting of three brain regions that orchestrate the AU value computation(40) (Figure 3H). The AU value signal, representing evolving anticipation while waiting, was encoded in the ventromedial prefrontal cortex (vmPFC)(40) (Figure 3I). The dopaminergic midbrain encoded an (anticipatory) RPE arising from information cues(40), consistent with previous findings about the anticipation of reward in MID tasks, which also have been associated with the traditional “wanting” circuit (known in the dichotomy of “wanting” vs “liking”)(12). The AU framework disentangles this phasic RPE signal from the evolving AU value signal, providing precise computational roles to the dopaminergic midbrain and the vmPFC.

Perhaps the most striking fMRI finding was the involvement of the hippocampus(40), an area rarely highlighted in previous studies using the MID tasks. In the AU task, participants’ hippocampal activity was significantly higher after receiving reward predictive cues than in the absence of such cues, and it was sustained over the entire 40-sec duration of reward anticipation(40). Further, hippocampal activity was functionally coupled to both the dopaminergic midbrain and the vmPFC, where the coupling strengths between these regions were predicted by the AU model(40). These findings suggest that the hippocampus plays a crucial role in boosting the AU. With the known role of the hippocampus in prospective thinking(38) and the anticipation of novel stimuli(36,154), the identified vmPFC–hippocampus–dopaminergic midbrain circuit may suggest AU value computation as an example of cognitive reward processing mediated by the hippocampus, which could involve prospection or imagination.

Recent studies have shown ramping dopamine signals as animals approach reward(158,159). While ramping signals are commonly observed across brain regions(160,161), their role remains debated. Some suggest these signals align with RL models, such as continuous temporal difference errors (159,162,163). While dopamine ramping is unclear in the AU task (146), its interpretation as a continuous prediction error implies the sustained boosting of anticipatory utility. Together with “negative ramping” (164), reconciling dopamine ramping with AU models may offer additional insights into the neural circuitry underlying anticipation.

Discussion

Anticipation of future experiences has a profound impact on mental health. In this paper, we presented a brief overview of altered anticipation across psychiatric conditions and the methodological tools used to study it, including self-report questionnaires, task-based neuroimaging, and computational modeling. Each approach has strengths and limitations—questionnaires are accessible but lack mechanistic insight, while MID tasks reliably probe neural correlates but may not fully capture behavioral and circuit dynamics. Emerging computational models, like the Anticipatory Utility (AU) framework, offer testable predictions but require further validation. We propose that integrating these methods provides the most effective strategy for understanding altered anticipation.

While reward anticipation has traditionally been linked to mesolimbic dopamine circuits (e.g., the wanting circuit), the hippocampus has also been implicated in MID tasks(165,166). This finding is consistent with the neural circuit underlying AU computation(40), and a growing body of evidence indicating the hippocampus’ involvement in value computation, reward processing(167,168), and prospection(38). Hippocampal alterations are common across schizophrenia(29), depression(28), anxiety disorders (66), addiction (73), and eating disorders (124). Integrated battery of anticipatory tasks (e.g., MID, NPU, AU tasks in fMRI experiments) and the tasks that target other cognitive functions (e.g., effort vs. reward task(169), multi-armed bandit tasks(170), delayed match-to-sample tasks(171), episodic memory tasks(172), and Raven’s progressive matrices(173)) may generate integrated insights into the role of hippocampus in alterations in anticipation and its relation to broader cognitive deficits across disorders. As the hippocampus is not a monolithic structure(41),targeted neuroimaging could clarify the differential roles of subregions associated with cognitive and emotional processing across alterations.

This integrated computational task battery could advance precision medicine through computational psychiatry (Figure 5). Digital frameworks, such as online experimental platforms and machine learning techniques, can facilitate this application. As described above, models such as the AU paradigm can capture participant’s unique anticipation computation using the model’s parameters and distinguish between different underlying neural mechanisms, such as changes in the hippocampus, dopaminergic midbrain, vmPFC, or connections between them. Although further validation is needed, it may help track treatment effectiveness of medications or cognitive therapies. For example, patients with anticipatory anhedonia (red cluster) may show a decreased PE boosting relative to healthy populations (green cluster) (Figure 5C), with interventions shifting model’s parameters over time (Figure 5E). Large transdiagnostic datasets could reveal distinct computational subgroups, enabling a machine learning-based system to personalize treatment recommendations.

Figure 5. A potential application of the computational framework toward precision psychiatry.

(A). Many participants will perform a battery of tasks online or in-person. (B). Computational models can be fit to behavioral data to estimate parameters that characterize each participant’s anticipatory process. (C). The participants’ data can be mapped onto the model’s parameter space, where they could be clustered according to their distinct computational phenotypes. (D). Participants may undergo interventions. (E). Repeated administrations of the task could capture changes over time and/or the impact of interventions in the computational parameter space, tracking the potential trajectories in the progression of psychiatric disorders and their treatments. The model with the dataset could also generate personalized predictions about treatment outcomes.

Most studies, including the Research Domain Criteria (RDoC)(174), treat positive (e.g., reward anticipation)(15,16,175) and negative anticipation (e.g., anxiety, fear, sustained threat)(125,176) as separate processes. However, in real life, these processes are almost impossible to disentangle. For example, fear of failing a test coexists with the hope of passing. These complex anticipatory processes play critical roles in psychiatric conditions but remain understudied.

The AU computational framework unifies positive and negative anticipation, capturing symptoms across conditions (Figure 4). Reduced positive anticipation could explain anticipatory anhedonia, while heightened negative anticipation could explain anxiety. Healthy emphasis on positive over negative anticipation may relate to optimism bias(177). A balance shift between them could capture the complex dynamics of substance use and eating disorders(120,121) (Figure 4C), while in bipolar disorder, it may reflect mood-dependent behavioral shifts (178). Yet, neural substrates of anticipation could still be distinct for the valence of the outcomes, and further studies are needed.

Goal-directed and habitual controls (or model-based and model-free reinforcement learning (179)) (105,180) could impact the anticipatory processes. The AU process may begin as a goal-directed, model-based process, e.g., the early phase of substance use or the initial stages of AN. However, over time, it may gradually become habitual, transforming into a process that does not require explicit goals(105,112,181).

Recent computational modeling (182) suggests that early life experiences of uncertain reward environments could lead to altered reward timing estimation, potentially generating anhedonia-like symptoms. Integrating AU with uncertain-timing (183) or partially observable semi-Markov models (184,185) could incorporate this aspect.

In sum, we suggest integrating varieties of methods that target altered anticipation, including self-report questionnaires, the established task-based approach, and novel computational frameworks, can be a powerful way to advance our understanding of altered anticipation and decision-making processes. The new computational approach has the potential to elucidate common and distinct computational mechanisms underlying complex processes across psychiatric symptoms, including anhedonia, anxiety, and craving offering a promising avenue to help identify new therapeutic targets.