Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Computational Psychiatry and Clinical Model Translation

Abstract. Computational psychiatry develops mathematical and computational models of the processes implicated in psychiatric symptoms and disorders. The problem of clinical translation is to establish when an experimentally interpretable model becomes a reliable, transportable, clinically useful instrument for understanding individual trajectories, evaluating interventions, or supporting treatment decisions. This critical narrative research monograph traces the field from reinforcement-learning and Bayesian models through transdiagnostic research, computational phenotyping, symptom-specific applications, reliability and recovery studies, prediction-model evaluation, and prospective clinical implementation. It distinguishes mechanistic explanation, statistical association, individual prognosis, treatment-effect prediction, and demonstrated clinical benefit; reviews limitations of task measurement, model identifiability, population transfer, and decision thresholds; and synthesizes operational requirements for responsible model translation. The manuscript reports a literature-based synthesis, not a preregistered systematic review, clinical trial, regulatory authorization, or newly validated patient-level tool.

Keywords: computational psychiatry; clinical translation; computational phenotyping; reinforcement learning; Bayesian inference; predictive processing; mechanistic modeling; psychiatric symptoms; model identifiability; parameter recovery; test–retest reliability; treatment prediction; clinical validation; transdiagnostic psychopathology; precision psychiatry; decision support

The Clinical Translation Problem

Computational psychiatry studies psychological and neurobiological dysfunction through explicit models of information processing, learning, inference, choice, and regulation. A model connects observed clinical or experimental phenomena with a specified process that produces predictions under stated assumptions. Clinical translation asks a more demanding question: what evidence establishes that a model can improve the description, prognosis, prevention, or treatment of mental disorder in a defined clinical population? A model that reproduces group differences can identify a plausible process while remaining insufficiently reliable to classify individual patients or guide an intervention. The distinction is central to the research program articulated by Montague and colleagues and the later integration of theory-driven and data-driven computational psychiatry by Huys, Maia, and Frank Montague et al. (2012) Huys et al. (2016).

Clinical translation can be represented as a sequence of evidential claims. Construct interpretation concerns what a latent parameter means; measurement validation establishes whether it can be recovered and estimated reproducibly; clinical association asks whether the measure relates to symptoms, functioning, or treatment history; predictive validation examines performance in future or independent patients; and utility evaluates whether decisions informed by the model improve patient-relevant outcomes. Success at an earlier stage does not automatically establish a later claim. A robust neural correlate, for example, can support a mechanistic hypothesis without offering incremental prognostic accuracy over routine clinical information. The forward-translation problem and the importance of prospective testing are developed explicitly in Yip and colleagues' clinical perspective Yip et al. (2023) Huys et al. (2011).

This monograph uses a critical narrative synthesis of original experiments, methodological assessments, clinical prediction investigations, and contemporary peer-reviewed reviews. It evaluates the scope of computational psychiatry rather than attempting a quantitative pooled effect estimate. The principal questions are how psychiatric constructs become computationally operationalized, which results survive psychometric and independent validation, how model-derived quantities relate to clinical interventions, and what governance is appropriate when a mathematical estimate affects patient care. The unit of analysis alternates deliberately between the model, the behavioral task, the person, the clinical decision, and the health-care institution because each possesses distinct validity criteria Huys et al. (2016) Yip et al. (2023) Moons et al. (2025).

Historical Formation of the Field

The foundations of computational psychiatry precede its consolidation as an explicitly named research field. Mathematical accounts of learning, formal theories of decision making, engineering-inspired control models, and quantitative neuroscience established techniques for linking a process hypothesis to observable performance. Experimental reward-learning investigations and computational neuroscience subsequently supplied models capable of connecting trial-wise behavior to neural activity. A pivotal human investigation by Daw and colleagues in 2011 employed a multistep decision paradigm to distinguish influences of prospective, model-based evaluation from habits of model-free reinforcement learning, while demonstrating that striatal prediction-error signals could reflect both kinds of computation. This was a substantive experimental challenge to overly simple one-process assignments Daw et al. (2011).

Montague, Dolan, Friston, and Dayan's 2012 account framed mental dysfunction as potentially aberrant computation across levels of organization; Wang and Krystal's 2014 synthesis extended the bridge between neural circuits and psychiatric explanations. Their work joined computational description to mechanistic hypotheses, including learning, neural dynamics, cortical regulation, and symptom expression. These accounts became influential partly because psychiatric diagnostic groups often contain people with heterogeneous trajectories and processes. The promise was a more precise functional description than a diagnostic label can supply on its own; the challenge was preserving the explanatory and measurement commitments of each proposed computational construct Montague et al. (2012) Wang and Krystal (2014).

The 2010 Research Domain Criteria initiative provided a complementary research framework for studying dimensions of functioning across diagnostic categories and levels of analysis. Computational parameters can contribute to such dimensional investigations when their meaning and measurement validity are established. During the following decade, larger digital recruitment cohorts, hierarchical fitting, machine learning, and clinical neuroimaging expanded the range of questions. By the 2020s, methodological work increasingly assessed reliability, practical task burden, external validation, and implementation. This history is best understood as an expansion of testable research approaches with unequal degrees of clinical maturity, rather than as a completed displacement of symptom-based diagnosis or clinician judgment Insel et al. (2010) Hedge et al. (2020) Yip et al. (2023).

Levels of Explanation and Clinical Constructs

A computational account can specify the objective of a system, an algorithm transforming information, or a physical implementation. In psychiatric research these levels intersect with clinical phenomenology, lived context, developmental experience, neural physiology, and social conditions. A parameter such as learning rate may summarize the weight assigned to recent feedback within a particular task, yet the same numerical pattern could result from fatigue, misunderstanding, variable attention, inconsistent preferences, or nonstationarity in the environment. Relating computational quantities to neurobiology therefore requires converging experiments and interventions; fitting a behavioral model alone does not uniquely identify a biological mechanism Huys et al. (2016) Wang and Krystal (2014).

The construct-to-model chain runs from a clinical phenomenon to an operational task, a generative model, a fitted parameter distribution, and an empirical validity argument. At every link, a distinct hypothesis is introduced. A depressive symptom questionnaire may index interest or pleasure, an effort task may capture willingness to expend work, and a reinforcement-learning model may estimate sensitivity to reward and expected value. These measures overlap in domain but are not interchangeable. The 2026 critical appraisal of computational models of anhedonia found variation in the operationalization of this multidimensional symptom and limited predictive validation, providing an instructive example of how theoretical plausibility can exceed demonstrated clinical utility Singh et al. (2026).

A clinically useful model must also define its target of inference. Estimating how a patient currently responds to rewards differs from explaining why a symptom emerged, forecasting future relapse, identifying a change-sensitive biomarker, and choosing between treatments. Each target requires specific sampling, outcome measurement, timing, and validation. Mechanistic interpretability is valuable because it yields hypotheses that can be experimentally discriminated, but predictive discrimination, calibration, and decision consequences must be evaluated separately. No single statistical score establishes explanation, diagnosis, treatment effect, and patient benefit together Huys et al. (2011) Collins et al. (2024) Moons et al. (2025).

Reinforcement Learning and Value-Based Decisions

Reinforcement-learning models represent how choices change with experience. In a common prediction-error formulation, an updated value is obtained by adding a learning-rate-weighted discrepancy between observed and predicted outcomes. Choice rules can translate learned values into action probabilities, with an inverse-temperature parameter regulating how strongly value differences govern selection. More elaborate accounts distinguish separate sensitivities to rewards and punishments, exploration strategies, volatility-sensitive learning, and transitions between model-free habits and model-based planning. These variables are formally meaningful only relative to the task, the assumed outcome structure, and competing models Daw et al. (2011) Montague et al. (2012).

The clinical attraction of reinforcement learning is its ability to separate processes that aggregate accuracy and mean reward conceal. Two participants may earn similar rewards because one updates beliefs rapidly while selecting inconsistently and another updates slowly but chooses consistently. Conversely, a low observed reward score can reflect perceptual confusion, motor slowness, diminished effort, altered sensitivity to loss, or changes in expectations. A candidate model should generate distinctive behavioral predictions and should be tested against alternative parameterizations rather than described as the mechanism merely because its estimated parameters correlate with symptom severity Palminteri et al. (2017) Huys et al. (2016).

Theoretical value and clinical discriminability must be distinguished in relation to frontostriatal functioning. The multistep experiment of Daw and colleagues demonstrated that human decisions involve interacting model-based and model-free influences and that measured striatal activity can express more than a simple model-free error signal. Psychiatric interpretations that assign a single cognitive deficit to a single circuit or disorder risk compressing this complexity prematurely. The appropriate translation is an experimentally specified claim about a population and a process, followed by tests of parameter stability, symptom specificity, and meaningful outcome prediction Daw et al. (2011) Gillan et al. (2016) Schaaf et al. (2024).

Bayesian Inference, Predictive Processing, and Perceptual Beliefs

Bayesian models describe inference under uncertainty by relating prior expectations to information furnished by observations. In a perceptual task, the influence of an existing belief depends on assumptions about sensory likelihood and the relative precision of available evidence. Such formulations are powerful because they can generate explicit predictions about how context, expectation, and signal reliability interact. In psychiatry, they have been applied to perceptual abnormalities, belief updating, psychotic experiences, and uncertainty-dependent decisions. The same formal language supports multiple distinct hypotheses, so an interpretation should identify which prior, likelihood, or precision term has been measured and manipulated Montague et al. (2012) Wang and Krystal (2014).

Powers, Mathys, and Corlett studied hallucination-like perceptual experiences using Pavlovian conditioning in which an auditory event became associated with a visual cue. Their 2017 experimental report linked conditioned hallucinations to greater weighting of perceptual priors in a model-based account. The study illustrates a productive route from a phenomenological clinical feature to a task with experimental manipulation and a quantitative hypothesis. Its evidence is specific to the paradigm, participants, and inferential comparisons employed; clinical use would require separate evidence concerning discrimination among conditions, temporal stability, and the benefits of acting on the resulting measure Powers et al. (2017).

Bayesian explanation can also be misused when a fitted latent variable becomes an all-purpose synonym for a psychiatric symptom. Different perceptual, attentional, and decision processes can yield related behavioral signatures, and measurement noise can distort their fitted precision. Testing competing generative models, manipulating prior cues and sensory reliability independently, examining parameter recoverability, and extending observations across tasks are therefore central requirements. Model comparison should aim to separate competing predictions that matter to the clinical construct, rather than reward a computational account for flexible retrospective accommodation of an observed symptom Palminteri et al. (2017) Huys et al. (2016) Moons et al. (2025).

Depression, Anhedonia, and Reward Learning

Depression is heterogeneous with respect to mood, reward experience, energy, psychomotor changes, memory, and social functioning. Computational work has sought to decompose commonly grouped phenomena into reward responsiveness, reward expectation, effort valuation, exploration, and choice consistency. The process description matters because an individual who anticipates little benefit from an activity may differ functionally from someone who experiences a reward but cannot mobilize sufficient effort to obtain it. Nevertheless, questionnaires, laboratory reinforcement tasks, and model parameters cover overlapping but distinct aspects of the clinical construct. The 2026 scoping review by Singh and colleagues appraised 37 studies of computational anhedonia models and identified substantial limits in predictive validity, particularly beyond cognitive-behavioral mappings Singh et al. (2026) Huys et al. (2016).

Computational accounts have also been investigated as candidate prognostic measures. Bansal and colleagues reported reinforcement-learning-related markers associated with later remission in a study of depression, illustrating the scientific value of prospective outcomes rather than diagnosis-only comparisons. Such findings require careful attention to sample selection, competing baseline predictors, treatment heterogeneity, and the definition of remission. Predicting that participants with a certain behavioral or neural characteristic will generally recover more often is an outcome-prognosis claim; predicting which of two treatments will produce better outcomes for a specified person is a different, treatment-interaction claim Bansal et al. (2025) Furukawa et al. (2020).

The central problem for depression translation is to prove that a measured computational dimension is reproducible, responsive to meaningful change, and clinically informative beyond accessible clinical measures. Models may become more compelling when they predict specific treatment-related mechanisms and when changes in the proposed mediator temporally precede symptom improvement. Such an argument would need repeated measurements, validated symptom instruments, controlled intervention comparisons, and sensitivity analyses for medication exposure and comorbid states. On present evidence, the strongest contribution of many computational reward paradigms is mechanistic refinement and hypothesis generation rather than routine individual prescription Singh et al. (2026) Yip et al. (2023) Moons et al. (2025).

Compulsivity, Habits, and Goal-Directed Control

Obsessive-compulsive symptoms and related forms of compulsivity have motivated models of competition between goal-directed planning and habitual action. Goal-directed behavior incorporates expected consequences and the current desirability of outcomes; habitual control can favor repeated actions supported by prior reinforcement. An experimentally measured tendency to persist after changing contingencies may provide evidence about one component of this balance. Gillan and colleagues reported differences in goal-directed and habit-related behavior in an obsessive-compulsive sample, connecting a formal action-control hypothesis to a clinical disorder. The experimental phenomenon should be interpreted through the particular task and group, including its demands on working memory, attention, and understanding Gillan et al. (2011) Daw et al. (2011).

A later transdiagnostic study by Gillan, Kosinski, Whelan, Phelps, and Daw examined goal-directed control and psychiatric symptom dimensions in two independent general-population samples. The association with a dimension comprising compulsive behavior and intrusive thought was more specific than many broad diagnostic comparisons had suggested. This supports investigating dimensional patterns of psychopathology and computational control together; it does not demonstrate that the same computational process produces every clinically observed compulsion. Correlation with a symptom factor also leaves open the influence of shared task requirements and the interpretation of the fitted computational quantity Gillan et al. (2016).

For clinical model translation, the crucial test is whether the computational assay contributes incremental prediction or intervention guidance when added to symptom history, functional impairment, and existing assessments. A proposed intervention targeting action-control balance should modify a prespecified model-relevant behavior and demonstrate patient-important benefit under a controlled design. If a computational parameter is statistically associated with compulsivity yet unstable at retest, or if its estimated meaning changes with model specification, treating it as a patient-level biomarker would be premature. The available literature supports an active research pathway from transdiagnostic process evidence toward clinical validation Gillan et al. (2016) Hedge et al. (2018) Yip et al. (2023).

Psychosis, Hallucinations, and Inference under Uncertainty

Psychotic experiences include disturbances of perception, belief formation, salience, and the appraisal of internally generated experiences. Computational approaches have addressed these phenomena through reinforcement learning, predictive processing, Bayesian perception, and neural network models. The central scientific advantage of an explicit model is its capacity to specify distinct sources of an apparently similar clinical presentation. Strong priors, altered sensory reliability, attentional selection, learning history, and decision criterion changes can generate different predictions during experimentally manipulated tasks. A patient description of an anomalous percept provides the phenomenon to explain, while the model identifies hypotheses about how available evidence is combined Wang and Krystal (2014) Huys et al. (2016).

The conditioned-hallucination paradigm investigated by Powers and colleagues linked repeated pairing of visual and auditory cues to subsequent reports of hearing a sound under conditions in which it was absent. Model-based analysis associated the phenomenon with stronger prior influence. That experimental result offers a mechanistically constrained candidate account, but replication across contexts and validation of clinical outcome relevance are separate requirements. It is especially important to distinguish an experimentally induced perceptual report from the full psychiatric phenomenology and social consequences of ongoing hallucinations. Diagnosis, risk assessment, and therapeutic planning require information that extends beyond an auditory conditioning parameter Powers et al. (2017).

Reliable translation would require prospective testing in heterogeneous psychosis-spectrum populations, including patients experiencing different symptom patterns, medication histories, and cognitive impairments. The relevant endpoints might include symptom persistence, functional recovery, treatment response, or the ability to track clinically meaningful changes across time. Prediction must be benchmarked against routine assessment and calibrated for the population in which it is applied. A behavioral or perceptual assay that works well in a laboratory setting may change when the patient is unwell, distracted, fatigued, or assessed in a different environment. This imposes a practical demand for multiple-site and repeated-measures evidence before individual clinical use Yip et al. (2023) Schaaf et al. (2024) Moons et al. (2025).

Anxiety, Threat Learning, and Uncertainty

Anxiety-related problems often involve anticipation of adverse outcomes, attentional priority for threat, avoidance, heightened uncertainty, and the persistence of threat expectations after circumstances change. Formal learning models can represent asymmetries between learning from positive and negative prediction errors, expectations about the volatility of environmental contingencies, or the subjective cost assigned to uncertain outcomes. Computational descriptions allow researchers to design experiments in which different mechanisms imply different choices across trials. Translating them requires a clear distinction between adaptive caution under real uncertainty and distressing or functionally impairing symptom patterns Montague et al. (2012) Huys et al. (2016).

Threat-learning tasks can vary dramatically in reinforcement schedules, stimuli, response requirements, and participant comprehension. A parameter identified as threat learning in one task may reflect memory capacity or attentional fluctuations in another; distinct neural processes may also produce observationally similar responses. A clinically relevant research design therefore separates cue perception, contingency learning, anticipated cost, exploration, and effort to avoid harm where the data permit. Simulation-based parameter recovery and experiment designs that create discriminating predictions can make this decomposition empirically credible. Claims about an individual's anxiety mechanism additionally require evidence of measurement stability and cross-task construct validity Palminteri et al. (2017) Hedge et al. (2020) Schaaf et al. (2024).

The eventual treatment question is whether an assay predicts a clinically consequential difference between evidence-based intervention options or tracks a hypothesized process when an intervention is administered. General improvement during treatment does not establish that the model-specific process caused that change; regression to the mean, spontaneous fluctuation, therapeutic alliance, practice effects, and measurement expectations may contribute. Designs with randomized allocation, credible comparators, repeated symptom observation, and prespecified computational mediators are essential for separating plausible model stories from evidence that could guide care. The translation standard is therefore a causal and decision-relevant question, not simply a better statistical fit Yip et al. (2023) Collins et al. (2024) Moons et al. (2025).

Addiction, Impulsivity, and Transdiagnostic Processes

Substance-related and addictive behaviors have long attracted reinforcement-learning, habit, valuation, and intertemporal choice models. These frameworks examine how immediate outcomes, delayed consequences, uncertainty, conditioned cues, and action repetition shape selection. A model can distinguish preference for immediate reinforcement from failures to represent downstream outcomes, and a learning rule may describe when new consequences cease to update behavior. The psychiatric challenge is that a similar behavioral pattern can arise from multiple sources: changing opportunities, withdrawal, social pressures, co-occurring symptoms, or stable decision preferences. Clinical interpretation has to integrate the environment within which choices are made Montague et al. (2012) Huys et al. (2016).

Transdiagnostic research investigates dimensions appearing across conventional diagnostic boundaries. The relationship between goal-directed control and a compulsive/intrusive-thought dimension in Gillan and colleagues' investigation illustrates the strategy. A measured dimension can align more closely with a computational process than a nominal diagnostic category, but the resulting association still needs testing in treatment-seeking populations. Clinical categories are used for diverse purposes, including communication, eligibility for services, and treatment planning; computational phenotyping contributes a distinct information layer whose added usefulness must be demonstrated rather than assumed Gillan et al. (2016) Insel et al. (2010).

For translation into addiction or other transdiagnostic settings, prediction horizons and consequences require explicit definition. Forecasting short-term task behavior is different from anticipating a clinically significant relapse or identifying which treatment approach improves retention and functioning. Clinical deployment also encounters changing substance exposure, stress, housing, social support, and access to care, creating context shifts that may invalidate parameters estimated elsewhere. Longitudinal data, external validation and comparison with ordinary clinical predictors are necessary, with evaluation centered on whether model-guided decisions alter outcomes in a favorable way Yip et al. (2023) Furukawa et al. (2020) FDA (2026).

Experimental Tasks and Observational Measurement

A computational psychiatric task is an assay only when the behavioral data reliably constrain a target process. Design begins with stimulus construction, instructions, reward or feedback schedules, timing, exclusion rules, participant burden, and the intended contrast between candidate mechanisms. Trial-by-trial choice, response latency, omission, confidence, and physiological signals may provide complementary observables; each also contains measurement errors and nuisance variation. The operational definition of a computational construct should explicitly identify which observable patterns distinguish its variation from changes in attention, perceptual ability, motor control, and comprehension. A model fitting a task that fails to evoke the intended process cannot recover clinically meaningful information by computational sophistication alone Huys et al. (2011) Hedge et al. (2020).

Task burden becomes decisive in psychiatric cohorts. Participants may experience fatigue, medication effects, discomfort, cognitive difficulties, or limits on available clinic time, while repeated measurement creates learning and practice effects. Long laboratory tasks can improve the information available for parameter estimation at the cost of lower feasibility and representativeness. A shortened assay requires new simulations and empirical validation because eliminating trials may change identifiability or encourage reliance on different strategies. The clinical goal favors the smallest informative experimental protocol whose task demands, measurement error, and patient burden have been independently assessed Yip et al. (2023) Schaaf et al. (2024).

Observational and ecological measurements broaden the available evidence by collecting information closer to ordinary settings, including repeated symptom ratings and contextual events. Their strengths depend on missing-data mechanisms, platform accessibility, consent, and the possibility that assessment itself changes behavior. Laboratory signals and field observations should be connected by explicit hypotheses about transfer, rather than pooled as though they measure the same latent attribute on equivalent scales. A model validated in a selected, technologically comfortable sample may fail in a clinical population with different age, educational background, comorbidities, or access to devices Moons et al. (2025) Collins et al. (2024).

Generative Model Specification and Estimation

A generative model defines a probability distribution over behavior or other observations conditional on latent variables and parameters. In a reinforcement-learning task, it might describe state values, a prediction-error updating rule, and a stochastic action-selection mechanism. In a perceptual task, it could describe how prior beliefs and uncertain sensory evidence determine responses. Parameter estimation may use maximum likelihood, Bayesian inference, or hierarchical modeling, each with assumptions about variability and uncertainty. Researchers should articulate the mathematical hypotheses and the range of plausible alternatives before estimating psychiatric group contrasts. Without that step, model flexibility can reproduce observed behavior while leaving psychological interpretation underdetermined Daw et al. (2011) Huys et al. (2016).

Hierarchical estimators can pool information across participants and stabilize noisy individual estimates, but they also introduce assumptions about the population distribution. A participant-level posterior estimate is partly informed by group-level assumptions, so apparent stability must be examined across estimation strategies, prior choices, and recruitment samples. Regularization reduces variation under specified conditions and can make out-of-sample prediction better, but a narrow posterior does not compensate for a misspecified behavioral mechanism. Clinical readers need to know whether uncertainty intervals reflect sampling variation, parameter correlation, model uncertainty, and the quality of data available for the individual Schaaf et al. (2024) Moons et al. (2025).

Model criticism begins with posterior predictive or simulation-based checks asking whether the fitted model reproduces informative features of behavior. It continues with comparisons against theoretically serious alternatives, interventions that distinguish their predictions, and sensitivity analyses. Palminteri and colleagues emphasize falsification because a complex model's capacity to fit data is not equivalent to the confirmation of the proposed mechanism. Psychiatric applications should report the amount and kind of evidence by which a specific mechanistic account outperforms plausible competitors and whether the distinction persists in independent data Palminteri et al. (2017) Huys et al. (2011).

Parameter Recovery, Identifiability, and Model Mimicry

Structural identifiability asks whether distinct parameter settings can produce the same predicted observations under the specified design. Practical identifiability asks whether the finite, noisy data actually permit useful estimation. A model with correlated learning-rate and choice-noise parameters may fit observed choices well even if the individual parameters cannot be separated. Recovery experiments simulate data from known parameter values, refit the model, and examine whether the estimates recover those values across a scientifically relevant range. Model-recovery tests additionally examine whether an analysis can distinguish candidate generating models. These are prerequisite scientific exercises for treating a fitted parameter as an interpretable clinical indicator Huys et al. (2011) Palminteri et al. (2017).

Recovery and reliability answer different questions. Simulation-based recovery establishes properties of the estimator under the simulated generative assumptions. Test–retest reliability examines the stability of measured individual differences over actual occasions. Good recovery can coexist with low retest reliability when participants genuinely change state, shift strategy, or perform under different circumstances. Conversely, highly stable parameters can reflect a reproducible measurement artifact or a rigid task structure. Translation requires both a recoverable target and empirical evidence that its temporal variation matches the clinical claim, distinguishing a trait biomarker from a potentially useful state-sensitive process measure Schaaf et al. (2024) Hedge et al. (2018).

Identifiability must be reconsidered whenever an experiment moves into a new setting. Shorter tasks, altered incentive structures, repeated exposure, symptom change, medication, and age differences can modify the information present in observations. Model mimicry is especially important when two algorithms predict nearly identical distributions on the available task but suggest contrasting therapeutic mechanisms. The solution is experimental design that elicits divergent predictions, such as changes in reward probabilities, transitions, information availability, or outcome devaluation. A clinical translation claim becomes stronger when model separation is driven by controlled data rather than post hoc labeling Daw et al. (2011) Palminteri et al. (2017) Yip et al. (2023).

Psychometric Reliability and Longitudinal Stability

The reliability paradox arises because tasks designed to produce robust average experimental effects can have little between-person variability and therefore weak reliability as measures of individual differences. Hedge, Powell, and Sumner demonstrated the general issue for cognitive tasks, and later computational psychiatry discussions emphasized that translating group-level paradigms requires a different psychometric objective. A task may reproducibly establish a population process while performing poorly in ranking people across occasions. Clinical prediction and treatment stratification depend on how reliably a person's measured differences correspond to stable or clinically relevant variation, so the intended usage must drive evaluation Hedge et al. (2018) Hedge et al. (2020).

Schaaf, Weidinger, Molleman, and van den Bos assessed reinforcement-learning parameters in online bandit and reversal-learning tasks administered about five weeks apart. Their independent samples comprised 69 and 47 participants; reported intraclass correlations varied substantially across parameters and tasks, with many low estimates. The investigators examined simulations and mood-related within-person variation, illustrating why weak test–retest agreement is not solved by simply fitting a more complex model. A parameter may be affected by stress, happiness, task-specific demands, and sampling variation in ways that matter for clinical interpretation. Reliability must therefore be reported per parameter, task and population rather than inherited from the general reputation of a method Schaaf et al. (2024).

A translational assay needs a stated measurement purpose: stable subtype identification, monitoring change, estimating a fluctuating process, or predicting a later outcome. Stable trait measurement generally calls for high agreement across appropriately chosen retest periods. A state-monitoring assay instead requires evidence that its variation tracks meaningful within-person change rather than uncontrolled noise or rehearsal. Repeated assessments, alternate task forms, measurement-invariance tests, and explicit modeling of person-by-occasion interactions are potential solutions. Their value should be established empirically rather than presumed from high internal consistency or fitting accuracy Enkavi et al. (2019) Schaaf et al. (2024) Yip et al. (2023).

Prediction, Calibration, and External Transport

Clinical prediction requires defining a future outcome, a time horizon, a target population, and a realistic decision context. A computational feature that correlates with current symptoms may contribute no prospective information beyond baseline severity. A predictive model should be compared with simple clinical baselines, its predictors measured before the outcome, and its entire analysis pipeline protected from information leakage. Internal cross-validation estimates performance under particular sampling assumptions; independent external evaluation tests whether a fitted model performs when people, sites, workflows, and data collection differ. These distinctions are equally relevant for interpretable computational parameters and larger machine-learning feature sets Collins et al. (2024) Moons et al. (2025).

Discrimination describes the ordering of risk across individuals; calibration concerns agreement between predicted probabilities and observed event frequencies. A well-discriminating model may be poorly calibrated after changes in prevalence or referral pathways. When a clinical threshold triggers an intervention, false alarms, missed cases, resource constraints, and treatment harms become part of the utility calculation. The reporting framework TRIPOD+AI requires transparent documentation of prediction studies, while PROBAST+AI supplies an updated method for evaluating development quality, evaluation bias, and applicability across conventional regression and artificial-intelligence methods. Neither guideline establishes clinical utility by itself Collins et al. (2024) Moons et al. (2025).

Clinical transport concerns more than statistical replication. The target setting may differ in diagnostic mix, language, access to care, medication use, symptom severity, age and cultural context. A computational marker learned from a highly selected volunteer cohort may require recalibration, adaptation, or a complete reassessment of its underlying construct. External evaluation should report subgroup-specific performance where data permit, uncertainty around estimates, and consequences of applying the model in each setting. Longitudinal changes in health systems and illness presentation create distribution shift over time, suggesting the need for monitored deployment rather than a permanent one-time validation Yip et al. (2023) Moons et al. (2025) FDA (2026).

From Laboratory Mechanisms to Clinical Endpoints

Translational progression begins with a mechanism that makes experimentally testable predictions and continues through measurement qualification, symptom relevance, patient-level validation, and outcome-oriented assessment. The clinical problem determines whether an assay is intended for screening, monitoring, prognosis, prediction of comparative treatment benefits, or evaluation of a specific intervention mechanism. These claims should be formulated separately because each requires a different design and decision standard. A parameter that distinguishes two research groups could be useful for studying the causes of symptoms while carrying little discriminative information within a clinically referred population. Forward translation remains relatively uncommon despite the growth of mechanistic insight Yip et al. (2023) Huys et al. (2016).

Outcome selection shapes the evidential result. A symptom scale, a categorical remission definition, functioning, subjective quality of life, treatment burden, and adverse events capture different dimensions of benefit. Clinical outcome measures should be defined before model evaluation and aligned with the length of time over which the model could reasonably influence the outcome. A one-session estimate of reward sensitivity cannot be assumed to predict a several-year illness course simply because its process interpretation sounds relevant. Appropriate longitudinal analysis should distinguish transient state changes, treatment exposure, attrition, disease progression, and regression toward the mean Bansal et al. (2025) Furukawa et al. (2020).

The scientific translation record should contain explicit intended-use statements, available clinical comparators, parameter recovery results, psychometric evaluation, prospective and external outcome evidence, subgroup performance, and the decision rule that would use a model output. It should also document circumstances in which the output should not be used. This makes it possible to evaluate precisely which link in the translation pathway is established and which remains hypothetical. A well-designed research tool can have enduring scientific value without passing the evidential threshold for a diagnostic product or individual treatment recommendation Yip et al. (2023) Collins et al. (2024) FDA (2026).

Prognostic Prediction and Treatment-Effect Heterogeneity

A prognosis model estimates an outcome under the patterns of care present in its data. A treatment-selection model must go further by estimating how a patient's outcome would differ between interventions. These are different estimands. A feature strongly associated with recovery may fail to predict a comparative treatment effect, because patients with favorable prognosis can recover under several strategies. Clinically meaningful personalization requires treatment contrasts derived from data in which the alternatives can be compared credibly, ideally through randomized treatment assignment with adequate representation of treatment-effect heterogeneity Furukawa et al. (2020) Yip et al. (2023).

Furukawa and colleagues reanalyzed a randomized clinical megatrial of second-line antidepressant strategies involving 1,544 participants, investigating whether multivariable prediction would meaningfully change recommendations relative to average treatment effects. The analysis produced differences for a minority of patients and emphasized the need for stronger predictive variables. This is an illuminating example of how an apparently precise person-specific decision may offer limited incremental benefit when treatment outcomes are similar for many candidates. It is not a demonstration that a computational learning parameter has independently established individualized antidepressant selection; the intervention-prediction problem requires its own evaluation Furukawa et al. (2020).

Evidence of a treatment-by-marker interaction in the derivation sample remains preliminary until evaluated under a credible independent protocol. Clinical performance should be measured by patient-relevant consequences of a decision policy rather than by a striking association, attractive risk score, or algorithmic complexity. A model's benefit can vanish when clinicians already use similar information, when a recommended treatment is inaccessible, or when predicted improvements are too small to justify adverse effects and costs. Prospective comparative effectiveness studies and explicit net-benefit analysis are therefore natural end points of translation Yip et al. (2023) Moons et al. (2025) Collins et al. (2024).

Intervention Targets and Mechanism-Sensitive Trials

Mechanism-oriented computational psychiatry can support intervention development by specifying a process that a therapy, drug, or behavioral training program is expected to alter. An intervention hypothesis should identify the model variable, the manipulation, the anticipated change in behavior, and the patient-important outcome. For example, increasing reward pursuit could result from stronger reward valuation, better prediction updating, improved effort allocation, or changes in environmental opportunity. An intervention that improves task reward totals does not identify which of these mechanisms changed unless the task and model support a discriminating test Huys et al. (2016) Yip et al. (2023).

A mechanism-sensitive trial requires repeated observations of the hypothesized mediator and outcomes at time points adequate for establishing temporal order. Randomization and comparison conditions address important alternative explanations, while preregistered models, blinded assessment where feasible, and intention-to-treat analysis help preserve interpretability. A computational variable may also respond to repeated test exposure or medication-related motor changes rather than the intended therapeutic process. Parameter recovery under simulated intervention effects and prospective testing of measurement invariance can clarify which observations would count as supporting or falsifying the causal mechanism Palminteri et al. (2017) Schaaf et al. (2024).

The evidence needed to guide clinical use is ultimately intervention-level. An assay could be highly reliable and strongly predictive but still fail to improve care because the recommended treatment is unavailable, produces adverse effects, or is no better than usual selection. Conversely, a model might prove useful for optimizing low-burden follow-up even without complete knowledge of the underlying neural mechanisms. The translational research question is whether a defined model-informed strategy produces better outcomes under realistic practice conditions. That question requires a prospective intervention evaluation, not only model-fit comparisons or retrospective clinical correlations Yip et al. (2023) FDA (2026) Moons et al. (2025).

Clinical Workflow and Decision-Support Implementation

A computational assay enters clinical work through specific processes: consent, task administration, data cleaning, model execution, result interpretation, communication with patients, and integration into subsequent decisions. Each stage can change the accuracy and meaning of its output. A tool requiring extensive specialist setup may be scientifically informative but impractical in a crowded service; a brief tool may be feasible while losing parameter identifiability. The intended user, time available, follow-up responsibility, missing-data rule, and escalation pathway should therefore be specified before implementation. A clinically interpretable report must show the validated purpose and uncertainty of the model rather than presenting latent parameters as self-evident diagnoses Hedge et al. (2020) Yip et al. (2023).

Human oversight has an epistemic function. Clinicians can detect situational information outside the training data, assess symptom meaning, and consider patient goals, comorbidities, medication and contextual constraints. But human review does not automatically correct a systematically biased algorithm; professional reliance can itself be shaped by presentation, automation bias, and workload. Prospective workflow studies should test whether users understand probabilities, parameter uncertainty, and appropriate limits, and whether model information materially improves rather than merely complicates decisions. Patient participation in the assessment of tolerability and usefulness is part of the translation problem Moons et al. (2025) Collins et al. (2024).

The U.S. Food and Drug Administration's January 2026 final guidance on clinical decision-support software distinguishes certain functions excluded from device regulation under specified statutory criteria from software functions that meet the definition of a medical device. Regulatory treatment depends on intended use, functionality and other facts rather than the presence of a computational model or a psychiatric label alone. This monograph describes research concepts and validation requirements; it does not supply regulated software or assert clearance or approval for any particular algorithm. A deployment plan must establish the applicable requirements before patient-facing use FDA (2026).

Fairness, Privacy, and the Artificial Era

Computational psychiatry combines highly sensitive clinical information with choices, symptom ratings, sometimes neuroimaging, and potentially repeated digital observations. The resulting datasets can encode vulnerabilities that demand confidentiality, proportional collection, informed participation, and meaningful control over secondary use. Psychiatric labels may also influence access to services and treatment if model outputs are misunderstood. Governance therefore concerns the full data lifecycle, including consent, storage, linkage, model development, external validation, correction of errors, and withdrawal arrangements where feasible. External validation needs attention to underrepresented populations, measurement equivalence, site differences, and patterns of missingness Moons et al. (2025) Collins et al. (2024).

Fairness is not guaranteed by a model's transparency or mathematical simplicity. A seemingly interpretable learning-rate measure can become inequitable if task instructions, language demands, motor controls, incentives or baseline outcome prevalence differ across groups. Performance should be examined in populations relevant to intended clinical deployment, with recognition that disaggregated analyses require adequate samples and intervals reflecting uncertainty. Design adaptations should avoid compromising construct identity, and recalibration should be assessed empirically. Independent governance and responsible reporting are parts of building trustworthy evidence rather than decorative steps after model development Moons et al. (2025) FDA (2026).

The Artificial Era creates new possibilities for adaptive assessments, multimodal data processing, and dynamic computational partnerships in research and clinical systems. These possibilities require scientifically explicit distinctions between models of human psychiatric behavior and the computational systems used to analyze or assist with that behavior. An artificial system capable of fitting an interpretable model is not thereby a validated clinician, and its apparent fluency does not establish mechanistic or clinical correctness. The appropriate research trajectory evaluates artificial analytic capabilities through controlled experiments, representative data, reproducible software, appropriate accountability, and prospective demonstrations of benefit for the humans receiving care Huys et al. (2016) Yip et al. (2023) FDA (2026).

Critical Appraisal of the Evidence Base

The evidence base of computational psychiatry is heterogeneous in design and strength. Controlled experiments establish that particular manipulations alter measurable behavior in a way that favors specified process explanations. Observational studies demonstrate associations between computed quantities and symptoms but are vulnerable to confounding and selection. Reliability studies test whether an assay can characterize people across relevant occasions. Prospective prediction studies assess incremental performance on future outcomes. Randomized decision-strategy evaluations address whether acting on computational predictions improves patient-relevant results. These forms of evidence should be read together because a persuasive theoretical story assembled from separate correlations does not become proof of an effective clinical intervention Huys et al. (2016) Yip et al. (2023) Moons et al. (2025).

An important source of uncertainty is selective reporting across the large space of possible tasks, mathematical models, preprocessing decisions, and symptom outcomes. Parameter naming can obscure differences between algorithms; two papers reporting a learning rate may have implemented different choice rules, parameter bounds, priors, and fitting routines. Large apparent group contrasts can also be sensitive to exclusion practices and unbalanced samples. Independent replication is especially important when a model is selected because it produces the expected psychiatric result. Researchers should preserve complete code, preregistered primary hypotheses when appropriate, simulated recovery results, alternative-model checks, and uncertainty estimates for all clinically relevant comparisons Palminteri et al. (2017) Hedge et al. (2018) Collins et al. (2024).

The empirical results illustrate both promise and constraints. Powers and colleagues provided an experimentally manipulated account of conditioned hallucinations, while Gillan and colleagues identified a transdiagnostic association between action control and compulsive symptoms. Schaaf and colleagues found that several common reinforcement-learning parameters were unreliable across test sessions under their examined conditions. A 2026 scoping review of computational anhedonia models found generally weak predictive validation despite promising cognitive-behavioral construct relationships. These findings support continued mechanistic investigation and focused assay improvement, and they place a substantial burden of independent evidence on claims about clinical prediction or routine individual treatment selection Powers et al. (2017) Gillan et al. (2016) Schaaf et al. (2024) Singh et al. (2026).

What Models Explain and What They Predict

A computational model is explanatory when its variables and transformations represent a plausible process with testable consequences. A model is predictive when it performs adequately on prespecified new outcomes under relevant conditions. Either property may exist without the other: a compact mechanistic theory can isolate a real group-level effect yet be too noisy for individual prognosis, while a high-dimensional prognostic model can produce useful risk estimates without revealing a single neurocognitive mechanism. In psychiatric translation, neither label should substitute for an evidential argument. Identified mechanisms need interventions or discriminating experiments; predictions need calibration, external evaluation, and a decision analysis Huys et al. (2016) Wang and Krystal (2014) Moons et al. (2025).

Computational phenotyping proposes that fitted process parameters can characterize clinically relevant individual variation. Its strongest version requires evidence that a parameter has a coherent operational meaning, can be estimated with adequate precision, behaves as expected under informative experimental manipulation, and varies across people in a reproducible way. Group-level correlations do not settle individual measurement accuracy. The reliability paradox cautions against inferring good between-person precision from large mean experimental effects. Hierarchical models and better task design may improve estimation, but their success must be observed in the intended population rather than inferred from general theoretical arguments Hedge et al. (2018) Schaaf et al. (2024) Enkavi et al. (2019).

A clinical model might also be useful without carrying a fully specified causal mechanism, for example if its prognosis is externally calibrated and decision-supported outcomes improve in a trial. Such uses still require exact language to distinguish a risk score from a causal treatment rule. Mechanistic interpretation is scientifically valuable because it proposes which intervention to test; empirical utility is clinically decisive when the model changes a care decision. Scientific synthesis should therefore report the degree of evidence for explanation, measurement, prediction and benefit as separate dimensions of the same research program Yip et al. (2023) Furukawa et al. (2020) FDA (2026).

A Translation-Focused Research Agenda

The next phase of computational psychiatry requires research designs organized around precisely defined clinical questions. For each proposed tool, investigators should prespecify the target symptom or functional outcome, intended population, decision horizon, assay, comparison with usual assessment, acceptable measurement error, and consequences of incorrect decisions. Simulations should demonstrate that the task distinguishes the relevant model mechanisms over plausible patient behavior. Pilot testing should estimate feasibility and patient burden, while reproducibility studies should test both between-person and within-person measurement under realistic clinical circumstances Huys et al. (2011) Hedge et al. (2020) Yip et al. (2023).

Multisite studies should then test transport across referral routes, demographic groups, language, comorbid conditions and treatment practices. Individual prognosis models require independent calibration and discrimination assessments, and treatment-selection models need credible comparisons of intervention effects. Where the proposed mechanism informs a new intervention, repeated computational measurement can contribute to mediation tests, but improvements in clinical outcomes remain a distinct endpoint. Reporting should conform to current prediction-model guidance when relevant, with risk-of-bias appraisal and explicit attention to data leakage, inappropriate flexibility and fairness Collins et al. (2024) Moons et al. (2025).

An effective scientific infrastructure includes interoperable task specifications, preregistered analysis plans where appropriate, open simulation code, versioned estimation software, shared definitions of clinical outcomes, and documentation of failed replication or transfer. Clinical partners and patients should participate from the beginning so that research targets correspond to feasible decisions and valued outcomes. Longitudinal cohorts and adaptive assessment systems can test whether model-derived measures capture changing illness states. The ultimate standard of progress is cumulative, independently reproducible evidence connecting a formal model to a scientifically meaningful construct and then to improved decisions or interventions Yip et al. (2023) Singh et al. (2026) FDA (2026).

Interdisciplinary Boundaries and Scientific Synthesis

Computational psychiatry intersects with mathematical psychology, cognitive neuroscience, psychometrics, clinical psychology, epidemiology, and the study of artificial computational systems. Mathematical psychology contributes formal description and model comparison; neuroscience supplies additional evidence concerning implementation; psychometrics tests measurement quality; clinical research defines meaningful outcomes and interventions; epidemiology clarifies populations, selection and transport; and computer science develops reproducible estimation and prediction methods. This division of labor matters because a scientific conclusion can be valid at one level without providing a warranted conclusion at another. Psychiatric explanation must remain sensitive to environments, social relations, and changing experiences as well as cognition and neural processes Huys et al. (2016) Wang and Krystal (2014).

The central synthesis of the reviewed literature is a layered account of evidence, not a new causal theory. Experimental tasks allow some computational distinctions to be studied with precision, and transdiagnostic associations offer valuable alternatives to diagnosis-only contrasts. Yet clinical prediction and treatment benefit require additional validation. Reliability, construct interpretation, parameter recoverability, model comparison, population transport and intervention utility are separate constraints whose combined satisfaction defines readiness for a specified intended use. The empirical diversity across depression, compulsivity and psychosis demonstrates why no single measurement strategy can be assumed to generalize across psychiatric phenomena Gillan et al. (2016) Powers et al. (2017) Singh et al. (2026) Schaaf et al. (2024).

For Psychology by Bogdanova, this Research Dossier occupies the space between general modeling theory and clinical use. The neighboring subjects of identifiability, computational model comparison, transdiagnostic functioning and algorithmic assessment retain their independent research questions. Here the distinct object is the translation pathway itself: how an explicit psychiatric model becomes a sufficiently validated tool to support patient-relevant inference. The Artificial Era extends the technological means of this inquiry, while empirical assessment remains the criterion for distinguishing promising modeling applications from demonstrated therapeutic or diagnostic value Huys et al. (2011) Yip et al. (2023) Collins et al. (2024).

Conclusion

Computational psychiatry has established a productive research language for describing learning, inference, valuation, action control and neural computation in relation to psychiatric phenomena. Original experiments and transdiagnostic investigations demonstrate that computational distinctions can illuminate processes hidden by aggregate scores or conventional diagnostic categories. The field's fundamental scientific strength lies in models that make assumptions explicit, generate discriminating predictions, and connect experimental manipulation to interpretable behavior. This research tradition has already shaped theoretical approaches to depression, compulsivity, psychosis, and multiple dimensions of psychological functioning Montague et al. (2012) Huys et al. (2016) Gillan et al. (2016) Powers et al. (2017).

The pathway from computational explanation to clinical translation is governed by additional evidential demands. A useful parameter must be identifiable and estimable under the task; the assay must exhibit appropriate reliability, construct validity and sensitivity to clinically meaningful variation; predictive models require independent evaluation and calibration; and decision-support strategies require evidence that acting on their recommendations benefits patients. These requirements respond to documented limits in the stability of common computational parameters and in the predictive validity of several symptom-focused modeling programs. They support a differentiated view of scientific maturity rather than an indiscriminate assignment of readiness to the entire field Schaaf et al. (2024) Singh et al. (2026) Moons et al. (2025).

The scientific horizon is the construction of prospectively tested relationships between psychiatric mechanisms, validated individual measurements and effective interventions. This requires collaboration among modelers, clinicians, experimental psychologists, psychometricians, patients, and researchers working on prediction and implementation. Computational psychiatry will contribute most effectively when formal models are evaluated against alternative explanations, transferred carefully across settings, and assessed by their capacity to improve outcomes that matter to people. Those conditions define the central research agenda for clinical model translation Yip et al. (2023) Collins et al. (2024) FDA (2026).

References

  • Bansal, V., McCurry, K. L., Lisinski, J., Kim, D.-Y., Goyal, S., Wang, J. M., Lee, J., Brown, V. M., LaConte, S. M., Casas, B., & Chiu, P. H. (2025). Reinforcement learning processes as forecasters of depression remission. Journal of Affective Disorders, 368, 829–837. https://doi.org/10.1016/j.jad.2024.09.066
  • Collins, G. S., Moons, K. G. M., Dhiman, P., Riley, R. D., Beam, A. L., Van Calster, B., Ghassemi, M., Liu, X., Reitsma, J. B., van Smeden, M., Boulesteix, A.-L., Camaradou, J. C., Celi, L. A., Denaxas, S., Denniston, A. K., Glocker, B., Golub, R. M., Harvey, H., Heinze, G., . . . Logullo, P. (2024). TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ, 385, e078378. https://doi.org/10.1136/bmj-2023-078378
  • Daw, N. D., Gershman, S. J., Seymour, B., Dayan, P., & Dolan, R. J. (2011). Model-based influences on humans’ choices and striatal prediction errors. Neuron, 69(6), 1204–1215. https://doi.org/10.1016/j.neuron.2011.02.027
  • Enkavi, A. Z., et al. (2019). Large-scale analysis of test–retest reliabilities of self-regulation measures. Proceedings of the National Academy of Sciences, 116(12), 5472–5477. https://doi.org/10.1073/pnas.1818430116
  • Furukawa, T. A., Debray, T. P. A., Akechi, T., Yamada, M., Kato, T., Seo, M., & Efthimiou, O. (2020). Can personalized treatment prediction improve the outcomes, compared with the group average approach, in a randomized trial? Developing and validating a multivariable prediction model in a pragmatic megatrial of acute treatment for major depression. Journal of Affective Disorders, 274, 690–697. https://doi.org/10.1016/j.jad.2020.05.141
  • Gillan, C. M., Kosinski, M., Whelan, R., Phelps, E. A., & Daw, N. D. (2016). Characterizing a psychiatric symptom dimension related to deficits in goal-directed control. eLife, 5, e11305. https://doi.org/10.7554/eLife.11305
  • Gillan, C. M., Papmeyer, M., Morein-Zamir, S., Sahakian, B. J., Fineberg, N. A., Robbins, T. W., & de Wit, S. (2011). Disruption in the balance between goal-directed behavior and habit learning in obsessive-compulsive disorder. American Journal of Psychiatry, 168(7), 718–726. https://doi.org/10.1176/appi.ajp.2011.10071062
  • Hedge, C., Bompas, A., & Sumner, P. (2020). Task reliability considerations in computational psychiatry. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 5(9), 837–839. https://doi.org/10.1016/j.bpsc.2020.05.004
  • Hedge, C., Powell, G., & Sumner, P. (2018). The reliability paradox: Why robust cognitive tasks do not produce reliable individual differences. Behavior Research Methods, 50(3), 1166–1186. https://doi.org/10.3758/s13428-017-0935-1
  • Huys, Q. J. M., Maia, T. V., & Frank, M. J. (2016). Computational psychiatry as a bridge from neuroscience to clinical applications. Nature Neuroscience, 19(3), 404–413. https://doi.org/10.1038/nn.4238
  • Huys, Q. J. M., Moutoussis, M., & Williams, J. (2011). Are computational models of any use to psychiatry? Neural Networks, 24(6), 544–551. https://doi.org/10.1016/j.neunet.2011.03.001
  • Insel, T., Cuthbert, B., Garvey, M., Heinssen, R., Pine, D. S., Quinn, K., Sanislow, C., & Wang, P. (2010). Research domain criteria (RDoC): Toward a new classification framework for research on mental disorders. American Journal of Psychiatry, 167(7), 748–751. https://doi.org/10.1176/appi.ajp.2010.09091379
  • Montague, P. R., Dolan, R. J., Friston, K. J., & Dayan, P. (2012). Computational psychiatry. Trends in Cognitive Sciences, 16(1), 72–80. https://doi.org/10.1016/j.tics.2011.11.018
  • Moons, K. G. M., et al. (2025). PROBAST+AI: An updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ, 388, e082505. https://doi.org/10.1136/bmj-2024-082505
  • Palminteri, S., Wyart, V., & Koechlin, E. (2017). The importance of falsification in computational cognitive modeling. Trends in Cognitive Sciences, 21(6), 425–433. https://doi.org/10.1016/j.tics.2017.03.011
  • Powers, A. R., Mathys, C., & Corlett, P. R. (2017). Pavlovian conditioning–induced hallucinations result from overweighting of perceptual priors. Science, 357(6351), 596–600. https://doi.org/10.1126/science.aan3458
  • Schaaf, J. V., Weidinger, L., Molleman, L., & van den Bos, W. (2024). Test–retest reliability of reinforcement learning parameters. Behavior Research Methods, 56(5), 4582–4599. https://doi.org/10.3758/s13428-023-02203-4
  • Singh, S., Cunningham, J. E. A., Uher, R., Becker, S., & Nunes, A. (2026). The conceptualization, measurement, and critical appraisal of computational models of anhedonia in depression. Neuroscience & Biobehavioral Reviews, 185, 106652. https://doi.org/10.1016/j.neubiorev.2026.106652
  • U.S. Food and Drug Administration. (2026). Clinical decision support software: Guidance for industry and Food and Drug Administration staff (January 2026). https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software
  • Wang, X.-J., & Krystal, J. H. (2014). Computational psychiatry. Neuron, 84(3), 638–654. https://doi.org/10.1016/j.neuron.2014.10.018
  • Yip, S. W., Barch, D. M., Chase, H. W., Flagel, S., Huys, Q. J. M., Konova, A. B., Montague, P. R., & Paulus, M. P. (2023). From computation to clinic. Biological Psychiatry: Global Open Science, 3(3), 319–328. https://doi.org/10.1016/j.bpsgos.2022.03.011