Computational Psychiatry

BackgroundPavlovian responding is a core component of behavior and can be measured via Pavlovian-instrumental transfer (PIT), where Pavlovian responses bias instrumental actions. Standard single-lever PIT paradigms, which assess responses using a single-choice option, cannot dissociate the contribution of model-free versus model-based reinforcement learning. While indirect evidence suggests a role for model-free responding in single-lever PIT, the contribution of model-based strategies is unclear. It also remains unknown whether internal cognitive states, such as mind wandering, impair specifically model-based but not model-free PIT, as is theoretically expected. MethodsWe developed a novel, trial-by-trial two-stage PIT paradigm designed to computationally dissociate model-free and model-based Pavlovian responding by leveraging probabilistic state transitions and trial-wise outcome predictions. After each two-stage Pavlovian learning trial, participants performed a single-lever PIT trial as well as a query trial of explicit value judgment. Detailed task instructions were provided to support potential model-based strategies. Computational modeling was used to quantify individual learning strategies. We assessed mind-wandering questionnaires and thought probes. ResultsAnalysis of query and PIT trials revealed trial-by-trial updating of outcome expectations based on probabilistic task structure, consistent with model-based Pavlovian responding. Behavioral responses during PIT were best explained by a computational model-based reinforcement learning model. In contrast, we found little evidence for model-free Pavlovian responding. Higher levels of mind wandering were associated with reduced model-based control but did not impact model-free indices. ConclusionWe introduce a novel single-lever PIT paradigm that enables fine-grained dissociation of model-free versus model-based Pavlovian response systems. Our findings provide evidence that single-lever PIT can operate through model-based mechanisms, challenging the assumption that single-lever PIT is predominantly model-free. Our findings also indicate that internal attentional states selectively modulate model-based PIT. Given the involvement of Pavlovian responding in numerous psychiatric conditions, our paradigm offers new avenues for understanding maladaptive behavior. Author SummaryOur daily actions are often influenced by cues like the smell of food or the sound of phone notifications that signal potential rewards or losses. These Pavlovian cues can shape our instrumental behavior even though their outcomes do not depend on what we do - a process known as Pavlovian-instrumental transfer (PIT). Here we study the computational learning mechanisms that underlie such PIT effects. While it is often assumed that Pavlovian responding follows simple, automatic rules without a cognitive model of cue consequences (i.e., model-free), evidence also shows a role for cognitive anticipations in Pavlovian responding (i.e., model-based). In this study, we extend this evidence by showing that PIT responding can be driven by flexible model-based learning. We designed a task to test whether participants use model-free versus model-based strategies to guide PIT, providing detailed task instructions. Using reinforcement learning models, we found that most participants used model-based learning when forming cue-outcome associations. Importantly, peoples attention mattered: when they were more distracted and doing mind wandering, they relied less on model-based strategies. Our findings suggest that Pavlovian learning is complex, flexible, and influenced by internal mental states, opening new windows to understand decision-making problems in mental health conditions like addiction.

Background: Early initiation of substance use is linked to later adverse outcomes, and risk factors come from multiple domains and are shared across substances. In our previous work, traditional time-to-event Cox models identified individual risk factors, but these models are not designed to jointly model multiple outcomes or capture complex non-linear relationships. Multi-task learning (MTL) can leverage shared structure across related outcomes to improve prediction and distinguish common versus substance-specific predictors. However, most MTL studies rely on baseline features and focus on single outcomes, which limits their ability to capture shared risk and temporal changes. Substance use initiation is a time-dependent process that unfolds during development and reflects changing exposures over time. Baseline-only models cannot capture these changes or represent risk dynamics. Discrete-time modeling provides a practical approach by estimating interval-level initiation risk and combining it into cumulative risk at the subject level. By integrating multi-task learning with dynamic modeling, it is possible to share information across outcomes while capturing how risk evolves over time, which may improve prediction performance. Methods: Using the Adolescent Brain Cognitive Development (ABCD) Study (release 5.1), we developed two complementary multi-task learning (MTL) frameworks to predict initiation of alcohol, nicotine, cannabis, and any substance use. A baseline MTL model predicted fixed- horizon (48-month) initiation using one record per participant, while a dynamic discrete-time MTL model incorporated longitudinal interval data to model time-varying risk. Both models used multi-domain environmental exposures, core covariates, and polygenic risk scores (PRS). Performance was evaluated on a held-out test set using AUROC, PR-AUC, and calibration metrics, and compared with single-task logistic regression (LR). Feature importance was assessed using permutation importance and compared with Cox proportional hazards models. Results: MTL showed comparable or improved performance relative to LR, with larger gains for low-prevalence outcomes (cannabis and nicotine). Incorporating longitudinal information led to consistent improvements across all outcomes. Dynamic models increased AUROC by +0.044 to +0.062 for MTL and +0.050 to +0.084 for LR, indicating that temporal information was the primary driver of performance gains. Feature importance analyses showed modest overlap across methods, with higher agreement between dynamic MTL and Cox models than static MTL. A small set of features, including externalizing behavior, parental monitoring, and developmental factors, were consistently identified across all approaches. Conclusions: Dynamic multi-task learning improves the prediction of substance use initiation by leveraging longitudinal structure and shared information across outcomes. While MTL provides additional gains, incorporating time-varying information is the dominant factor for improving performance. Combining baseline and dynamic frameworks offers a comprehensive strategy for identifying robust risk factors and modeling adolescent substance use initiation.

Deficits in inhibitory control are common across a wide range of psychiatric disorders and are closely linked to symptom severity, including emotional dysregulation, anxiety, substance misuse, and self-harm, making them an appealing target for intervention. Cognitive training offers a low-cost, scalable, and non-invasive strategy to strengthen inhibitory control; however, most existing paradigms target only a single facet of inhibition and rarely account for environmental influences, such as affective context. To address these gaps, we developed a computerized inhibitory control training paradigm to simultaneously engage three components of inhibition: preemptive, proactive, and reactive, while embedding trials within positive and negative affective contexts to assess the impact of emotional stimuli. Across two online experiments, participants completed the GAMBIT task in one session (Experiment 1, N = 300) or repeated over three sessions (Experiment 2, N = 65). The task included No-Go trials to train preemptive inhibition, stop-signal trials for reactive inhibition, and stop-signal anticipation trials to train proactive inhibition. Affective images of differing valence were presented as background stimuli to evaluate their impact on inhibitory performance. In Experiment 1, participants showed higher accuracy on No-Go versus reference Go trials ({beta}=1.45, SE=0.09, p<.001), confirming successful manipulation of preemptive inhibition. Reaction times were slower during anticipation trials across two different conditions ({beta}=0.16, SE=0.04, p<.001; {beta} = 0.07, SE = 0.04, p = 0.047), consistent with proactive slowing when anticipating a potential stop signal. Additionally, positive affective images ({beta} = 0.10, SE= 0.009, p < 0.001) further slowed RTs, indicating emotional interference with proactive control. In Experiment 2, the pattern of higher No-Go accuracy was replicated ({beta} = 0.91, SE = 0.11, p < .001) and accuracy generally improved over sessions ({beta} = 0.38, SE = 0.06, p < .001). In anticipation trials, RTs become shorter across sessions (session 2: {beta} = -0.25, SE = 0.06, p < .001; session 3: {beta} = -0.45, SE = 0.06, p < .001), reflecting practice-related gains, and SSRTs decreased over time (F(2,56) = 6.26, p = .004), consistent with enhanced reactive inhibition. Proactive inhibition was modulated by affective images, with both negative ({beta} = 0.04, SE = 0.02, p = .039) and positive ({beta} = 0.16, SE = 0.02, p < .001) affective images associated with slower RTs. Participants also reported reductions in self-assessed temper control by the last session (W = 25.5, p = .007, q = .037, d = -0.51) and usability ratings were high (all means [&ge;] 3.87/5). Together, these findings show that this paradigm recruits multiple forms of inhibitory control and yields training-related improvements in both performance and affective outcomes. This provides preliminary validation of a scalable, fully online inhibitory control training tool targeting multiple dissociable inhibitory processes within affective contexts. The approach holds promise as an accessible transdiagnostic intervention to support symptom improvement across psychiatric disorders, with future work needed to evaluate clinical efficacy in patient populations.

Inferring the true cause of noise--distinguishing between volatility (environmental change) and stochasticity (outcome randomness)--is essential for learning in noisy environments. While most studies rely on binary outcomes, previous models are designed for continuous outcome and use ad hoc approximations to handle binary data, introducing theoretical inconsistencies and interpretational issues. Here, we develop a normative framework for inferring the causes of noise from binary feedback that remains faithful to the discrete nature of the generative process and underlying statistical structure. First, we establish a generative model using a state space approach tailored for binary outcomes and derive the corresponding hidden Markov model inference procedure. Second, we introduce a computational model combining the hidden Markov model with particle filtering to simultaneously infer volatility and stochasticity from binary outcomes. Third, we validate predictions through a 2x2 probabilistic reversal learning task with human participants, systematically manipulating both noise parameters. Results show that participants adjust their learning rates consistent with model predictions, increasing learning rates under volatile conditions and decreasing them under high stochasticity. Our theoretical and experimental results offer a principled approach for dissociating volatility and stochasticity from binary outcomes, providing insights into learning processes relevant to typical cognition and psychiatric conditions.

In 2024, approximately 30% of U.S. adolescents reported having consumed alcohol at least once in their lifetime, with about 25% of these individuals engaging in binge drinking. Adolescent alcohol use is associated with neurodevelopmental impairments, elevated risk of later alcohol use, and mental health disorders. These findings underscore the importance of identifying the variables driving adolescent alcohol use and leveraging them for early identification and targeted intervention. Previous studies have typically developed machine-learning classification models that use neuroimaging data in combination with limited clinical measurements. Neuroimaging data are expensive and difficult to obtain at scale, whereas clinical measures are more practical for large-scale screening due to their low cost and widespread accessibility. However, clinical-only approaches for alcohol drinking classification remain largely underexplored. Furthermore, prior studies have often focused on adults, limiting generalizability to the broader adolescent population. Additionally, confounding factors such as age and substance use, which are strongly correlated with alcohol consumption, have often been inadequately addressed, potentially inflating classification performance. Finally, class imbalance remains a persistent challenge, with prior attempts yielding only limited improvements. To address these limitations, we propose FocalTab, a framework that integrates TabPFN with focal loss for robust generalization and effective mitigation of class imbalance. The approach also incorporates an initial preprocessing step to remove confounding factors to account for age and substance-use. We compare FocalTab against state-of-the-art methods across different variable selections and dataset settings. FocalTab achieves the highest accuracy (84.3%) and specificity (80.0%) in the most stringent setting, in which both age and substance use variables were excluded, whereas competing models drop to near-chance specificity (12-24%). We further applied SHapley Additive exPlanations (SHAP) analysis to identify key clinical predictors of drinker classification, supporting enhanced screening and early intervention.

Inhibition is a core cognitive control function whose competence is distributed across the population, with more extreme impairments in psychiatric conditions such as attention deficit hyperactivity disorder (ADHD). The Stop Signal Task (SST) is a widely used paradigm for assessing this ability. However, conventional formalizations of SST performance, such as the independent race model, rely on assumptions that are frequently violated in modern experimental designs. Furthermore, the typical focus is on fitting mean reaction times, overlooking trial-by-trial dynamics. To address these limitations, we model the SST as a partially observable Markov decision process. This framework characterizes inhibitory control through distinct components: noisy perceptual inference regarding stimuli, and optimal control balanced against potential costs. To assess the ability of the model to capture the distribution of inhibitory capacities, we fit it to data from the large Adolescent Brain Cognitive Development (ABCD) study baseline cohort (N = 5,114). To do this, we adapted Simulation-Based Inference with a transformer-based encoder. This architecture learns compact, sequence-aware embeddings from raw behavioral data. These embeddings enable amortized inference of individual-level parameter posteriors in an efficient and reliable end-to-end manner, as confirmed by extensive validation. We identified distinct computational phenotypes associated with ADHD traits. Children with higher ADHD scores exhibited greater directional imprecision, a diminished intrinsic penalty for inhibition failures, and a more deterministic response style. Notably, the learned embedding space reveals a continuous manifold where children with the higher ADHD scores are heterogeneously distributed, rather than forming distinct disorder clusters. This indicates that similar clinical traits can emerge from diverse combinations of computational mechanisms, supporting a dimensional perspective on neurodiversity. Our framework can be extended to a broader range of cognitive tasks, offering a scalable solution for fitting complex models to large-scale behavioral data. Author summaryInhibitory control is essential for adjusting thoughts and behavior and is often impaired in conditions like ADHD. Traditional models of the Stop Signal Task (SST) often oversimplify the complex decision-making involved. We formalized these cognitive processes using a more biologically grounded framework (POMDP). This approach separates perceptual processing from control adjustments and remains valid in diverse experimental designs where traditional models fail. To apply the model at scale, we developed a specialized machine learning approach (TeSBI). This allowed us to efficiently reverse-engineer individual cognitive profiles. Applying it to the ABCD dataset (which includes more than 5,000 children), we found that higher ADHD scores are linked to specific computational deficits: noisy sensory processing, a lack of concern for errors, and a deterministic response style. Crucially, children with higher ADHD scores did not form a single disorder cluster but displayed diverse cognitive combinations, supporting a dimensional view of neurodiversity. Our results show that our model effectively captures complex inhibition mechanisms. By combining theory-driven cognitive modeling with scalable data-driven inference, this framework enables the precise analysis of large-scale behavioral datasets. This paves the way for more personalized approaches in computational psychiatry by recognizing the heterogeneity within clinical traits.

BackgroundPsychotherapy is proven efficacious for the treatment of alcohol use disorder (AUD). However, the patient-perceived importance of its effect is not fully appreciated in the evidence base. The smallest worthwhile difference (SWD) represents the smallest beneficial effect of an intervention that patients deem worthwhile in exchange for the harms, expenses, and inconveniences associated with the intervention, and facilitates the interpretation of patient perceived worthiness of an intervention. MethodsThe proposed study will estimate the SWD of NIAAA recommended psychotherapies for AUD treatment with English-speaking American respondents aged 18 and older. Primary participants will be recruited using the Prolific research crowdsourcing site. The SWD will be estimated using the Benefit-Harm Trade-off Method, presenting survey respondents with variable, hypothetical magnitudes of psychotherapy outcomes to find the smallest acceptable effect over a natural remission alternative. The overall average SWD, and subgroup distributions by participant AUD treatment experiences and AUD symptomology will be described. Secondary findings will estimate the smallest recommendable risk difference for AUD psychotherapy from providers and criminal justice professionals. Expected ResultsWe expect to find an estimate of the SWD for AUD psychotherapy. Further, we expect that the SWD will vary between clinical subgroups based on AUD symptomology and treatment experiences. We expect differences in SWDs between the general population and those of providers and criminal justice professionals. Findings from this project will inform the treatment decision process about psychotherapy during the clinical consultation for people with AUD.

Pediatric bipolar disorder is challenging to diagnose accurately due to symptom heterogeneity. More standardized and data-driven approaches are needed to enhance diagnostic reliability. We evaluated a clinical decision tool (nomogram), statistical methods (logistic regression, LASSO), machine learning (support vector machine, random forest, k-nearest neighbors, extreme gradient boosting), and deep learning model (multilayer perceptron) for pediatric bipolar disorder prediction across two datasets collected in academic (N=550) and community (N=511) clinical settings. We compared three modeling strategies: cross-dataset validation, cross-dataset with interaction terms, and mixed-dataset. We assessed model performance using discrimination ability, calibration, and predictor importance ranking. In the baseline cross-dataset approach, all models showed good internal discrimination in the academic dataset; but external discrimination in the community dataset substantially declined. Interaction-enhanced models slightly improved internal discrimination but not external performance or calibration. Recalibration prominently improved cross-dataset calibration without compromising discrimination, indicating that transportability problems were largely driven by probability scaling. Models trained on mixed datasets exhibited much stronger external discrimination and calibration. Across models and training strategies, family risk and PGBI-10M were consistently ranked as the most important predictors. Predictive models for pediatric bipolar disorder showed strong internal performance but limited cross-setting generalizability due to dataset shift and miscalibration. Increasing model complexity did not improve external performance, whereas training on pooled data substantially improved both discrimination and calibration. Findings suggest that sampling diversity, rather than model complexity, is more valuable for developing clinically useful and generalizable psychiatric prediction models, underscoring the importance of open and collaborative datasets.

We propose a computational instantiation of three cognitive stages from the Dot-Linear- Network (DLN) framework, grounded in a compression-efficiency thesis. DLN stages are characterized as graph-structured belief-dependency representations used to evaluate options: Dot as no persistent belief graph (reactive policies with negligible internal state), Linear as a null graph over option beliefs (K independent option estimates with no information sharing), and Network as shared latent structure (a bipartite factor graph in which F latent factors connect to K options), augmented by a temporal exposure state and an explicit structural learning cycle (hypothesis [-&gt;] test [-&gt;] update/expand). We distinguish two compression targets--option-factor structure (shared components in expected outcomes) and stakes-factor structure (shared drivers of consequence-bearing exposures)-- whose intersection yields jointly efficient actions that simultaneously improve expected outcomes and marginal exposure impact. In a bandit-like simulation (100 seeds, K [isin] { 20, 50, 100, 200 }, F =5), Network policies dominate Linear policies in cost-adjusted utility at large K, with the empirical crossover occurring much earlier than an analytic cost-only prediction (K* = F + cmeta/cparam), revealing that the advantage is primarily statistical (shrinkage-like estimation gains from factor pooling) rather than purely computational. Under stakes, all non-DLN agents--including Linear-Plus agents with identical factor structure and Network-standard agents with hierarchical Bayesian learning--collapse due to unmodeled cumulative exposure, while Network-DLN maintains positive utility. Within-stage consistency tests (two algorithmically distinct agents per stage) confirm that the collapse pattern is determined by representational topology, not algorithmic choice. These results evaluate internal consistency of a DLN-to-computation mapping under explicit assumptions; they do not validate a developmental theory in humans.

Humans have a remarkable ability to act efficiently and accurately in familiar situations while remaining flexible in novel circumstances. Nonparametric contextual inference has been proposed as a computational principle that can model how agents achieve flexible yet stable behaviour in dynamic and possibly unknown environments. However, it remains an open question how humans learn, deploy and reuse stable contextual task representations so efficiently. To address this question, we propose the nonparametric Bayesian Contextual Control (NP-BCC) model, which integrates nonparametric contextual learning with two well-established cognitive mechanisms: repetition-based automatisation and schema-like prior knowledge. These two mechanisms are assumed to support behavioural stability and facilitate novel task acquisition. Simulations in dynamic multi-armed bandit tasks of increasing difficulty illustrate how the NP-BCC can acquire and reuse contextual task representations, with the proposed mechanisms operating in the intended, functionally meaningful manner. Specifically, we show via simulations that automatisation not only enhances task performance but also stabilizes contextual inference and structure learning, while structured prior knowledge accelerates the acquisition of novel contexts. We discuss the implications of our findings for computational accounts of adaptive behaviour and contextual learning, and outline directions for future empirical work, including investigations of context-dependent behavioural dysregulation relevant to conditions such as substance use disorders. Author summaryPeople are very good at repeating well-learned actions in familiar situations, but they can also quickly adjust their behaviour when circumstances change. How the brain balances stability and flexibility is still not fully understood. There is growing evidence that the brain organizes experience into different "contexts", which are mental representations of encountered situations. Computational models based on this idea can in principle reproduce flexible behaviour, but they often become unstable in complex environments. To improve stability, we borrow two simple strategies from everyday human behaviour. First, people tend to repeat actions that have worked well before. Second, when facing something new, they often reuse strategies from similar past situations. Using simulations, we show that combining these strategies with context-based learning produces more reliable behaviour in the model. Prior experience helps the model understand new situations more quickly, while repeated actions help stabilise behaviour once a situation becomes familiar. Taken together, our findings show how such mechanisms can give rise to both flexible and stable behaviour in the model.

Purpose: Suicide and self-harm are major public health concerns characterized by substantial clinical and psychosocial heterogeneity. While latent class analysis has been used to identify subgroups of people with suicidal behavior, the extent to which such population-level phenotyping complements explainable artificial intelligence-based classification models remain unclear. Methods: We applied latent class analysis to a cross-sectional, publicly available dataset of 1000 individuals presenting with self-harm and suicide-related behaviors at Colombo South Teaching Hospital, Kalubowila, Sri Lanka. Sociodemographic, psychosocial, and clinical variables were used to identify latent subgroups. Class characteristics and suicide prevalence were examined and compared with variable importance patterns reported in a previously published explainable artificial intelligence (XAI)-based suicide classification study using the same dataset. Results: Four latent classes were identified. Two classes exhibited very high suicide prevalence (91.2% [95% CI: 87.7-93.8] and 99.0% [95% CI: 96.4-99.7]), whereas two classes showed low prevalence (<1%). The two high-prevalence classes differed markedly in lifetime psychiatric hospitalization history, with one class showing a 100% prevalence of prior hospitalization and the other substantially lower hospitalization rates. These patterns partially aligned with, and extended beyond, variable importance findings from the XAI-based model. Conclusion: Latent class analysis identified distinct subgroups with substantially different suicide prevalence and clinical profiles, underscoring the heterogeneity of individuals presenting with self-harm. Comparison with XAI-based suicide classification model findings suggest that unsupervised phenotyping and supervised classification provide complementary perspectives, offering population-level context that may enhance the interpretability of suicide assessment frameworks. Keywords: suicide; self-harm; latent class analysis; explainable artificial intelligence; machine learning

General distress cuts across psychiatric symptom domains, yet its computational correlates remain poorly defined. We examined whether drift rate--a core parameter indexing the efficiency of evidence accumulation--is more strongly associated with general distress than with domain-specific symptoms. In a cross-sectional online sample of 441 adults from the general population, participants completed a perceptual and value-based decision-making task, symptom assessments, and cognitive testing. Drift rates were estimated using hierarchical drift-diffusion modeling. Individuals with severe symptom elevations showed robust reductions in drift rate, particularly for value-based decisions. Mixed-effects models demonstrated that general distress, indexed by the Positive Symptom Distress Index, was more strongly associated with value-based than perceptual drift rate, even after accounting for all symptom domains. Value-based drift rate also explained variance in general distress beyond that accounted for by elevated symptoms across domains and selectively attenuated associations with somatization and paranoid symptoms. These findings suggest that value-based evidence accumulation captures a transdiagnostic component of distress-related impairment that is not reducible to symptom burden alone.

BackgroundMillions of people use language models to discuss mental health concerns, including suicidal ideation, but limited frameworks exist for evaluating whether these systems respond safely. Benchmarking, the practice of administering standardized assessments to language models, offers direct parallels to clinical competency evaluation, yet few clinicians are involved in designing, validating, or interpreting these assessments. AimsTo introduce mental health professionals to benchmarking language models by administering a validated clinical instrument and demonstrating how configuration decisions, measurement limitations, and scoring context affect result interpretation. MethodWe administered the Suicide Intervention Response Inventory (SIRI-2) programmatically to nine commercially available language models from three providers. Each item was presented 60 times per model (three prompt variants x two temperature settings x 10 repetitions), yielding 27,000 model responses compared against point-in-time expert consensus. ResultsTotal scores ranged from 19.5 to 84.0 (expert panel baseline: 32.5). Prompt design alone shifted individual model scores by as much as the difference between trained and untrained human groups. The best performing model approached the instruments measurement floor. All nine models consistently overrated clinically inappropriate responses that sounded supportive. ConclusionsA single benchmark score can support markedly different claims depending on the assumed standard of clinical behavior, the instruments remaining measurement range, and the configuration that produced the result. The skills required to make these distinctions must become core competencies. Benchmark results are increasingly utilized to support claims about mental health safety that may not be accurate, making it necessary to close the gap between clinical measurement and AI. Plain Language SummaryAI chatbots like ChatGPT, Claude, and Gemini are increasingly used by millions of people to discuss mental health problems, including thoughts of suicide. To assess whether these systems handle such conversations safely, researchers give them standardized tests called benchmarks and compare their answers to those of human experts. These scores are already used to argue AI systems are ready for clinical use. This study gave a well-established test of suicide response skills to nine AI models from three major companies under varying conditions. We changed how much instruction the AI received and how much randomness was built into its responses, then measured whether the scores changed. The same AI model could score like a trained crisis counselor under one set of conditions and like an untrained undergraduate under another, depending on choices the person running the test made. Every model also made the same kind of mistake: responses that sounded warm and caring were rated as appropriate, even when experts had judged them to be clinically problematic. The highest-scoring model performed so well that the test could no longer measure whether it was truly skilled or had simply exceeded the tests range. These findings show that a single score can be misleading without knowing how the test was run, whether it can still distinguish strong from weak performance, and whether it matches what the AI is used for. Mental health professionals routinely make these judgments about clinical assessments and are well positioned to bring that expertise to AI evaluation.

A widely used framework for studying the computational mechanisms of decision making is the Drift Diffusion Model (DDM). To account for the presence of both fast and slow errors in empirical data, the DDM incorporates across-trial variability in parameters such as the drift rate and the starting point. Although these variability parameters enable the model to reproduce both fast and slow errors, they rely on the assumption that over trials each parameter is independently sampled. As a result, the DDM effectively predicts that errors-- whether fast or slow--occur randomly over time. However, in empirical data this assumption is violated, as error responses are often temporally clustered. To address this limitation, we introduce the autocorrelated DDM, in which trial-to-trial fluctuations in drift rate, starting point, and boundary evolve according to first-order autoregressive (AR1) processes. Using simulations, we demonstrate that, unlike the across-trial variability DDM, the autocorrelated DDM naturally accounts for temporal clustering of errors. We further show that model parameters can be reliably recovered using Amortized Bayesian Inference, even with as few as 500 trials. Finally, fits to empirical data indicate that the autocorrelated DDM provides the best account of error clustering, highlighting that computational parameters fluctuate over time, despite typically being estimated as fixed across trials.

Background: Large language models (LLMs) are increasingly used in mental health contexts, yet their detection of suicidal ideation is inconsistent, raising patient safety concerns. Objective: To evaluate whether an independent safety monitoring system improves detection of suicide risk compared with native LLM safeguards. Methods: We conducted a cross-sectional evaluation using 224 paired suicide-related clinical vignettes presented in a single-turn format under two conditions (with and without structured clinical information). Native LLM safeguard responses were compared with an independent supervisory safety architecture with asynchronous monitoring. The primary outcome was detection of suicide risk requiring intervention. Results: The supervisory system detected suicide risk in 205 of 224 evaluations (91.5%) versus 41 of 224 (18.3%) for native LLM safeguards. Among 168 discordant evaluations, 166 favored the supervisory system and 2 favored the LLM (matched odds ratio {approx}83.0). Both systems detected risk in 39 evaluations, and neither in 17. Detection was highest in scenarios with explicit suicidal ideation and lower in more ambiguous presentations. Conclusions: Native LLM safeguards frequently failed to detect suicide risk in this structured evaluation. An independent monitoring approach substantially improved detection, supporting the role of external safety systems in high-risk mental health applications of LLMs.

BackgroundSuicidal ideation (SI) fluctuates over time, yet traditional static risk factors poorly align with its dynamics over time. Understanding dynamic symptom patterns may advance knowledge of the temporal interplay between SI and co-occurring symptoms in adults with depressive and anxiety disorders. Materials and methodsWe analyzed six waves (at baseline, and after 2, 4, 6, 9, and 13 years of follow-up) of the Netherlands Study of Depression and Anxiety (NESDA; n = 305, mean age 40.8 years, 62% female) in participants with any SI fluctuation over time. Variables included depressive, anxiety, mastery, and worry symptoms. Dynamic Time Warping (DTW) quantified within-person temporal alignment between SI and other symptoms, and an undirected network and forestplot visualized co-fluctuations. Analyses were stratified by age-groups and sex. ResultsOver the years, SI co-fluctuated most strongly with affective and anhedonic depressive symptoms, including sad mood, low capacity for pleasure, low general interest, pessimism, quality of mood, and decreased appetite. Select anxiety (terrified/afraid) and worry (overwhelming worries) items also aligned with SI, whereas mastery items did not. Patterns were broadly consistent across age and gender subgroups. Networks indicated that SI is part of a cluster of depressogenic symptoms but bridges to acute fear and persistent worry. ConclusionsSI is a dynamic phenomenon closely linked to specific depressive, anxiety, and worry symptoms. Interventions targeting mood instability, anhedonia, and uncontrollable worry, combined with real-time monitoring, may improve personalized suicide prevention. DTW provides a framework to identify long-term temporally proximal symptom patterns.

AimsDue to the multifaceted nature of "impulsivity", its measurement remains fragmented. Here, we developed the Risky Social Choices task to provide evidence for its validity and reliability, while testing the hypothesis that impaired access to implicit knowledge of negative long-term consequences is of distinct importance for "impulsive" decision-making in a general population sample. MethodsForty participants chose whether to engage in risk-taking behaviors, which combined web-based AI-generated videos with narrated hypothetical scenarios and measured worries related to negative long-term consequences, approach-related motivation for short-term rewards, response time to and accuracy of recognizing degraded auditory prime words denoting negative long-term consequences. ResultsA pre-registered multi-step regression model was constructed with worry, motivation, response time and accuracy as predictors and percentage of risky choices as the outcome. Among all predictors, only prime word recognition accuracy was significantly negatively associated with risky choices, confirming our hypothesis of the role of reduced implicit access to negative long-term consequences in risk-taking decisions. In contrast, approach-related motivation for rewards was the only predictor significantly positively related to percentage of risky choices. DiscussionAs predicted, the negative association between risky choices and implicit access to negative long-term consequences supports its role as a distinct aspect of "impulsivity". The novel task successfully captured this aspect, paving the way for a more precise neurocognitive characterization of clinical conditions where "impulsivity" plays a key role. The findings unveil the importance of implicit social sequential knowledge for impulsivity in neurotypical populations, so far only investigated in patients with brain lesions.

The Balloon Analogue Risk Task (BART) is widely used to assess risk-taking and impulsivity, yet existing computational models struggle to unify sequential and prior evaluation strategies or fully capture uncertainty-driven information-seeking behavior. To address this, we introduce a novel computational framework grounded in the Active Inference Framework (AIF), which conceptualizes behavior as the minimization of expected free energy. Model comparisons demonstrate that AIF-based models statistically outperform existing benchmarks. Furthermore, we applied this framework to investigate impulsivity in women with Premenstrual Syndrome (PMS). Our model revealed that the PMS group exhibited significantly higher values in inverse precision of policy ({beta}0) and the phase difference of this parameter was only observed in PMS group. This suggests that high {beta}0 serves as a robust computational marker, reflecting both the trait impulsivity inherent in PMS and its state-like exacerbation across the menstrual cycle. Lastly, our findings indicate that impulsivity in PMS manifests not as a learning deficit, but as heightened sensitivity to trial-by-trial sequential evaluation at the expense of stable, pre-planned prior policies. This framework provides a neurobiologically plausible and mechanistically granular understanding of risk-taking, offering new avenues for computational psychiatry.

BackgroundSuicidal ideation (SI) is a major global concern, yet its dynamic interplay with other symptoms remains poorly understood. ObjectiveTo identify symptoms that co-fluctuate with or temporally precede SI to improve warning signal detection and intervention. MethodsLongitudinal data from three Dutch psychiatric cohorts with lifetime internalizing disorders (16 waves from April 2020 until February 2022) were collected during the COVID-19 pandemic. We analyzed depressive, happiness, anxiety, loneliness, worry symptoms, and COVID-19-specific items only in those participants with SI fluctuations. Dynamic Time Warping (DTW) quantified within-person similarity between symptom trajectories and SI, and results were aggregated at group level. FindingsThe 307 participants (mean age 44.8 years; 61.6% female) showed increasing SI over time (p < .001). SI aligned with four depressive symptoms (i.e., sad mood, low self-esteem, low interest, and reduced happiness), two anxiety-related symptoms (i.e., fear of losing control, faintness), feeling abandoned, and overwhelming worrying. In directed DTW analysis, sad mood, hypersomnia, worrying about projects, and numbness/tingling showed significant temporal precedence before SI. ConclusionSI is embedded in a broad symptom network beyond depression. These results underscore the value of time-sensitive, idiographic monitoring using tools like DTW to capture the person-specific temporal pathways through which SI emerges and intensifies. Clinical implicationsThis study suggests a core group of affective, cognitive, and interpersonal symptoms that could serve as informative signals for evaluating changes in SI and may represent actionable targets for intervention. Summary BoxO_ST_ABSWhat is already known on this topic?C_ST_ABSO_LISuicidal ideation (SI) is a dynamic phenomenon, yet traditional research often relies on static, group-level averages that do not capture individual fluctuations. C_LIO_LIWhile SI is linked to depression, it can emerge independently through complex interactions with other affective and interpersonal states C_LI What this study adds?O_LIThis study identifies a set of affective, cognitive, and interpersonal symptoms, sad mood, overwhelming worry, and feelings of abandonment, that significantly co-fluctuate with SI over weeks and months. Additionally four specific "leading" symptoms, sad mood, hypersomnia, worrying about projects, and somatic numbness, were found that precede increases in SI. C_LI How this study might affect research, practice or policy?O_LIThe identified co-fluctuations and precursors serve as informative "(early) warning signals" that can improve individual risk stratification and clinical monitoring and may represent targets for intervention. C_LIO_LIThe results support a shift toward network-based models in suicidology, emphasizing the need for time-sensitive monitoring to capture the complex and dynamic nature of suicidality. C_LI

ImportanceLarge language models (LLMs) have demonstrated diagnostic potential in several medical specialties, but their application to psychiatry - where diagnosis relies heavily on clinical judgment, narrative interpretation, and reasoning under uncertainty - remains insufficiently evaluated. ObjectiveTo evaluate diagnostic accuracy and clinician-judged reasoning quality of multiple large language models using psychiatric case vignettes. DesignMixed-methods evaluation study of diagnostic accuracy across four LLMs using 196 psychiatric case vignettes (135 published and 61 novel). Clinical reasoning quality was evaluated on a randomly selected subset of 30 vignettes using structured clinician ratings along two reasoning dimensions. The highest-performing model was illustratively compared with psychiatry trainees on the same subset. Diagnostic correctness for the full vignette set was assessed by a separate adjudicator LLM. SettingPublicly available model interfaces, December 2025. ParticipantsFive board-certified psychiatrists evaluated model-generated clinical reasoning. Two psychiatry residents served as the illustrative human comparison. Main Outcomes and MeasuresDiagnostic accuracy and clinician-rated clinical reasoning quality. Diagnostic accuracy was assessed using top-1 accuracy, top-5 accuracy, recall@5, and mean reciprocal rank based on ranked lists of five differential diagnoses per vignette. Clinical reasoning quality was assessed using two 5-point Likert scales adapted from the American Council of Graduate Medical Education Psychiatry Residency Milestones, evaluating data extraction and diagnostic reasoning. ResultsAcross 196 psychiatric case vignettes, Claude Opus 4.5 (Anthropic) achieved the highest diagnostic accuracy (top-1 accuracy, 0.638; top-5 accuracy, 0.801; recall@5, 0.731; mean reciprocal rank, 0.710) and clinician-rated reasoning scores. Higher clinician-rated diagnostic reasoning quality was strongly associated with diagnostic correctness in mixed-effects logistic regression analyses ({beta} = 1.80; p < 0.001), corresponding to an approximately six-fold increase in odds of a correct diagnosis per 1-point increase in reasoning score. In an illustrative comparison, diagnostic accuracy of Claude Opus 4.5 fell within the range observed for psychiatry trainees. Conclusions and RelevanceLLMs demonstrated high diagnostic accuracy and generated clinical reasoning that clinicians judged to be largely coherent and safe. Diagnostic reasoning quality was more strongly associated with diagnostic correctness than data extraction quality, underscoring the importance of evaluating reasoning alongside accuracy when assessing LLMs for clinical decision support in psychiatry. Key PointsO_ST_ABSQuestionC_ST_ABSCan multiple large language models accurately diagnose psychiatric conditions and generate diagnostic reasoning that clinicians judge as coherent, safe, and clinically meaningful? FindingsAcross 196 psychiatric case vignettes, four large language models demonstrated high diagnostic accuracy. In a clinician-evaluated subset of 30 vignettes, model diagnostic accuracy fell within the range observed for psychiatry residents. Clinicians judged model-generated diagnostic reasoning to be largely coherent and safe. Higher clinician-rated reasoning quality was strongly associated with diagnostic correctness, independent of data extraction quality. MeaningEvaluating diagnostic reasoning, in addition to accuracy, may be important when assessing large language models for potential clinical decision support in psychiatry.