Conference Agenda

Over the last decades, computational theories have transformed our understanding of the mind in both health and psychiatric disorders. Active Inference unifies perception, action, and learning under the mantle of Bayesian inference, where they emerge resulting from an agent’s drive to minimize free energy. This symposium brings together researchers who have applied Active Inference to behavioral and neural data in diverse mental health conditions, offering novel insights into underlying mechanisms.

First, Ryan Smith will discuss research on mechanisms of interoceptive Bayesian inference in affective and substance use disorders, shedding light on deficits in how individuals perceive their own bodily state. Riccardo Maromotti will present insights on illness awareness in Alzheimer's patients using a novel active inference model for the emotional Stroop task. Elisabeth Sterner will focus on the use of Active Inference to explain aberrant action selection in psychosis, providing insights into cognitive and behavioral disruptions seen in conditions like schizophrenia. Anna-Lena Eckert will present findings from a study of generative models in interpersonal behavior in depression, where aberrant model parameters lead to maladaptive social decisions in a trust game context. Finally, Andreea Diaconescu will present transdiagnostic findings using Active Inference to model suicide risk across psychiatric conditions.

Together, these presentations illustrate how Active Inference offers a cohesive computational approach to modeling diverse clinical phenomena, with implications for diagnosis and treatment. The symposium will foster discussions on the future directions and the potential of applying the Active Inference framework in clinical research, as well as current challenges and limitations.

How the brain detects and interprets signals from within the body – a process known as interoception – may play an important role in generating subjective feelings and contribute to psychiatric disorders. While interoception has received growing attention from researchers in recent years, the precise computational mechanisms through which the brain processes interoceptive signals remain unclear. In this talk, I will present recent computational modelling studies we have performed to better characterize these mechanisms across cardiac and gastrointestinal interoception. First, I will describe newly replicated results of modeling heartbeat perception as Bayesian inference, which suggest that subjective estimates of the reliability (precision) of cardiac signals may be less flexible in multiple psychiatric patient samples (depression, anxiety, substance use, and eating disorders) relative to healthy participants. Second, I will describe results of modeling gastrointestinal (GI) perception in a similar manner during EEG recording. As hypothesized, results show that individual differences in prior expectations, and in subjective estimates of the precision of GI signals, have inhibitory and excitatory influences on neural responses, respectively. Data also suggest stronger prior expectations against feeling stomach sensations in eating disorders, and asymmetric learning rates that maintain this bias. Overall, these results provide evidence for neurocomputational mechanisms of brain-body interactions across multiple interoceptive channels. They may also highlight novel mechanistic treatment targets that could be evaluated in future clinical studies.

Active inference is a theoretical framework that integrates perception and action, modeling behavior as an inferential process where actions are probabilistically selected based on prior knowledge and sensory input. In this talk, I will present an active inference model of the color-word Stroop task to investigate how voluntary mental effort influences cognitive control. Twenty healthy young adults performed the Stroop task under two conditions: exerting maximal effort or responding as relaxed as possible. Their behavior was analyzed using a two-layer Partially Observable Markov Decision Process (POMDP) that estimated two key parameters: each participant’s bias toward reading words (vs. reporting colors) and their motivation for accurate performance. Results indicated that voluntary effort selectively increased motivation for correct responses without altering habitual biases, suggesting that effort primarily enhances internal motivation rather than directly modulating habitual behaviors.

In the second part of the talk, I will extend this modeling approach to Anosognosia, a neurodegenerative condition characterized by a lack of illness awareness. This phenomenon can be examined using the Emotional Stroop task, another color-naming paradigm that measures interference from emotionally charged words (neutral, negative, and disease-related). Here, the POMDP model incorporates parameters for the salience of negative and disease-related words, allowing us to explain increased reaction times for disease-related words as a potential marker of implicit disease awareness. These findings have critical clinical implications, as impairments in effort regulation and illness awareness are common in neurodegenerative disorders such as Alzheimer’s disease.

To navigate their environment, humans need to build an internal model of the world to interpret ambiguous inputs. Inaccurate models, as suggested to be the case for individuals with psychosis, can disturb optimal action selection. Recent computational models, such as active inference, have emphasized the importance of action selection as a key part of the inferential process. Based on an active inference framework, we investigated the use of previous knowledge and belief precision in an action-based task, given that alterations in these parameters have been linked to the development of psychotic symptoms. We also assessed whether task performance and modelling parameters could classify patients and controls.

Methods: 23 individuals with an at-risk mental state, 26 patients with first-episode psychosis and 31 controls completed a probabilistic task in which action choice (go/no-go) was dissociated from outcome valence (gain or loss). We evaluated group differences in performance and active inference model parameters and performed receiver operating characteristic (ROC) analyses to assess group classification.

Results: We found reduced overall performance in patients with psychosis. Active inference modelling revealed that patients showed increased forgetting, reduced confidence in policy selection and less optimal general choice behaviour, with poorer action–state associations. Importantly, ROC analysis showed fair-to-good classification performance for all groups, when combining modelling parameters and performance measures.

Conclusion: Active inference modelling of this task provides further explanation for dysfunctional mechanisms underlying decision-making in psychosis and may be relevant for future research on the development of biomarkers for early identification of psychosis.

Social interactions are computationally challenging due to several volatile, unobservable factors at play. Interpersonal difficulties are common in mental disorders, but formal models of interpersonal functioning in psychiatry remain limited. The Active Inference (ActInf) framework puts inference over hidden states at the core of perception, action and learning and provides a suitable framework to investigate interpersonal decision-making in health and disorder.

For this study, N=56 controls and N=47 outpatients with depression played a Trust Game, where they can keep or invest a small monetary amount. If they invest, the amount is tripled, but they play against a partner who can decide to share the win evenly, or to defect. The game context was cooperative (80% returned investments) or hostile (20% returned). Participants were randomized to the game conditions.

Across contexts, controls earned higher rewards (M=30.82) than patients (M=29.86, p<0.05). To model decision-making, we fit a POMDP generative model with ActInf to participants' behavior, obtaining estimates for subject-specific parameters (i.e. matrices A, B, C, D, and epistemic drive ε). On average, controls showed a higher preference for keeping their money (Controls C=0.89, Patients C=0.77), and group-level behavior was best described by an increased epistemic drive in patients (Controls ε=0.51, patients ε=0.67). Across groups, we find a correlation between prosociality and epistemic drive ε (r=0.25, p=0.01). We are currently investigating parameter recoverability, and the accuracy of diagnostic classifiers trained on generative model parameters. These findings suggest that ActInf, coupled with behavioral data, can provide computational-level insights into altered social cognition in depression.

Suicide remains a public health concern, demanding mechanistic insights to improve early detection and personalized interventions. Building on an active inference framework, we formalize how hopelessness interacts with Pavlovian biases to drive maladaptive approach and avoidance strategies in suicidal individuals. Our model proposes four computational perturbations that heighten suicide risk: increased learning from aversive outcomes, reduced belief decay following unexpected events, heightened stress sensitivity, and decreased controllability of stressors. Drawing on neurobiological evidence of interactions between noradrenergic and cholinergic systems, we relate these perturbations to brain systems governing aversive learning and decision-making. We validate this framework using an Avoid/Escape Go/No-go paradigm in two clinical cohorts: 129 veterans with post-traumatic stress disorder subdivided into suicidal and non-suicidal groups, and 50 individuals with major depression exhibiting varying suicide risk levels. Hierarchical Bayesian analyses identified a Reinforcement Learning model with stable Pavlovian context biases and a forgetting factor as the best explanation for choice behavior at the population level, although an active inference model captured individual behavior well. Correlation analyses revealed substantial overlap between parameters, indicating that aversive sensitivity, go biases, and forgetting map onto prior preferences, action selection precision, and belief decay thresholds. Post-hoc subgrouping uncovered distinct suicidality subtypes—one marked by impulsive, stress-driven lose-switch tendencies and another by more planful, yet maladaptive, learning linked to increased hopelessness. Taken together, these results indicate stable Pavlovian biases in aversive contexts, amplified by recency effects, entrench patterns in suicidal groups. Moreover, active inference frameworks via prior preferences complement reinforcement approaches to refine understanding of suicidality.