Alterations in reinforcement learning (RL) have been linked to various forms of psychopathology in childhood and adolescence. In the LICA study, we investigated learning processes in children and adolescents using a Probabilistic Reversal Learning paradigm in a sample comprising healthy controls and clinical groups (affective, anxiety, and hyperkinetic disorders). Cross-sectional findings revealed associations between computationally derived RL parameters and symptom dimensions, including reduced learning from positive prediction errors in anhedonia and hyperactivity, as well as altered uncertainty-directed exploration in relation to fear (Falck et al., 2025).

The present follow-up study reassesses the same participants, now aged 10–20 years, approximately two years later. This longitudinal design allows us to examine to what extent changes in reinforcement learning processes are coupled with changes in psychopathological symptoms over time.

RL parameters are estimated using computational modelling and linked to dimensional measures of psychopathology and general well-being derived from self- and parental-reports. We further test whether RL parameters at baseline predict later symptom change, whether symptom levels predict changes in RL, and whether changes in both domains are coupled over time. Moreover, we explore whether symptom network structures differ across time points using network analysis and network comparisons test.

By integrating computational modelling with a longitudinal, transdiagnostic approach, this study aims to clarify whether alterations in RL represent risk factors associated with symptom progression. These findings may contribute to a better understanding of developmental pathways in psychopathology and inform more personalized approaches to diagnosis and intervention.

Negative expectations are increasingly recognized as relevant determinants of treatment outcome in anxiety disorders, yet the mechanisms through which they influence therapeutic change remain incompletely understood. In particular, it is unclear how generalized negative expectations translate into impaired learning during exposure therapy and ultimately poorer clinical outcomes.

In this talk, we present data from the large multicenter PROTECT-AD study examining a mechanistic pathway linking generalized negative expectations, treatment-specific expectations, and learning processes during exposure. Building on hierarchical models of expectation and predictive processing, we conceptualize generalized negative expectations as higher-order priors that bias more proximal beliefs about treatment success and shape subsequent learning dynamics.

Using sequential mediation models, we show that generalized negative expectations bias treatment-specific expectations, which in turn attenuate learning during exposure, consistent with a sequential mechanism of change.

Converging neuroimaging findings further indicate altered activation in the mid-cingulate cortex and bilateral insula/operculum in response to the CS+ compared to CS- in a fear extinction paradigm, suggesting that negative expectations modulate salience and cognitive control network dynamics relevant for reinforcement learning.

Together, these findings support a mechanistic account in which negative expectations impair exposure-based learning via altered expectancy updating, highlighting learning rate as a key parameter and negative expectations as potential target for personalized intervention.

A dynamical systems (DS) perspective has become increasingly central to psychology, psychiatry, and neuroscience, offering a principled mathematical language for capturing how brain activity and behavior evolve and change over time. DS models can forecast future trajectories, illuminate computational mechanisms, and identify critical transitions such as bifurcations or tipping points. Recent advances in machine learning have dramatically expanded our ability to learn flexible, nonlinear DS models directly from empirical data, without relying on hand-crafted mechanistic assumptions.

Yet the domains of psychology and psychiatry pose specific challenges that go beyond what standard ML-based DS approaches were designed to handle. In this talk, I present a number of recent methodological advances tailored to these challenges. First, I address the integration of multimodal and non-Gaussian data streams, such as ecological momentary assessment (EMA) time series and neural spike trains, into a unified DS modelling framework. Second, I discuss hierarchical inference approaches that learn shared dynamical structure across heterogeneous patient populations, enabling the identification of common mechanisms despite substantial inter-individual variability. Third, I discuss a novel class of DS models called almost-linear RNNs (AL-RNNs), which constrain nonlinearity to a small number of units, yielding sparsely nonlinear connectivity models that naturally connect to the widespread use of linear connectivity methods in neuroscience and psychopathology, while retaining the capacity to capture switching dynamics and state-dependent transitions.

Together, these contributions offer a flexible toolkit for reconstructing the nonlinear dynamics underlying neural and behavioral data.

We re-analyzed data (N = 345) from a multicenter randomized controlled trial on panic disorder with agoraphobia using network analytic approaches. Although exposure-based cognitive-behavioral therapy (CBT) is effective, a substantial proportion of patients does not sufficiently benefit, and mechanisms differentiating responders from non-responders remain poorly understood. Network theory offers a complementary perspective by conceptualizing symptoms and related processes as dynamically interacting elements that may sustain each other through vicious cycles. Within this framework, differences in network connectivity or density may relate to treatment response, with denser systems reflecting more persistent patterns.

We examined associations between markers of defensive reactivity in a controlled agoraphobic context and clinical indicators (e.g., self-reported panic severity, agoraphobic avoidance). The study included a standardized exposure-based CBT protocol with assessments at three time points (pre-, mid-, post-treatment), capturing both clinical outcomes and defensive reactivity during a behavioral activation test. Exploratory network analyses compared responders and non-responders in associations among physiological (e.g., heart rate), subjective (e.g., symptom intensity), and clinical indicators.

Across multiple estimation approaches, networks of non-responders appeared denser, indicating stronger and more interconnected associations. In line with network theory, this pattern suggests more tightly coupled systems that may be more easily activated and more resistant to change, potentially contributing to poorer treatment response. In contrast, responders tended to show weaker associations between markers of defensive reactivity and general panic symptoms.

These findings highlight the potential of network approaches to identify mechanisms underlying treatment response. Implications for developing prediction models of treatment outcome will be discussed.

The placebo effect has been shown in pain analgesia and depression, but evidence concerning the computational and neural underpinnings remains scarce. We present our recent work on the antidepressant placebo effect and its relationship with presumably dopamine-mediated reinforcement learning.
First, our findings show increased learning rates after induction of antidepressant treatment expectations, suggesting an expectation-driven increase in reward sensitivity. Second, pharmacological dopamine blockade impeded the expectation-driven learning rate increase, suggesting involvement of dopamine in the antidepressant placebo effect. Last, reinforcement learning parameters were significantly linked to EEG signatures, such as the P300 and RewP ERPs, and representational dynamicity.
We assumed that positive treatment expectations involve dopamine, which in turn increases reward sensitivity, thereby improving mood in the long run. Previously, we could show that positive treatment expectations and reward uncertainty increase reinforcement learning capabilities, while the link to dopamine was not assessed. In our latest pre-registered study, we therefore tested a sample of 297 healthy participants undergoing a 6-hour treatment in a crossed 2x2 between-subjects design, manipulating treatment expectations (high vs. low) and midbrain dopamine signalling (placebo vs. sulpiride 400mg pill). We collected (i) behavioural choice data to model reinforcement learning parameters, and (ii) EEG data to capture neural signatures related to dopamine and learning.
Taken together, our findings indicate an expectation-driven increase in cognitive flexibility, which may be used therapeutically to alleviate depressive symptoms, if these findings hold in a clinical sample as well.