Reconceptualizing Anxiety and Depression within the Free Energy Principle and Active Inference Framework


Reconceptualizing Anxiety and Depression: Insights from the Free Energy Principle and Active Inference Framework

The free energy principle and active inference framework offer a computational perspective for understanding psychiatric disorders, framing them as disorders of inference. This framework provides a unified explanation of how maladaptive beliefs and perceptions can lead to conditions like anxiety and depression by disrupting the brain’s ability to minimize uncertainty. For anxiety, traditional models have emphasized cognitive biases and learned responses, but the free energy principle suggests that anxiety stems from a persistent mismatch between expected and actual outcomes, leading to irreducible uncertainty. Similarly, in depression, the framework explains rumination as a failure in the brain’s generative models to effectively reduce uncertainty, leading to persistent negative thought patterns and maladaptive problem-solving strategies.

Anxiety

Traditional models of generalized anxiety suggest that false beliefs and biases lead to heightened and sustained anxiety responses. Early behavioral models focused on conditioning, where anxiety is learned in specific situations, causing anticipatory anxiety upon re-entering similar contexts. Pioneers like Beck and Ellis identified cognitive biases, such as catastrophizing and dualistic thinking, that distort perceptions and exacerbate anxiety. These biases result in a feedback loop where anxiety reinforces itself through misinterpretation of bodily sensations, as seen in Clark’s model of panic attacks. While cognitive-behavioral therapy (CBT) has been effective in clinical practice, it faces criticism because anxiety symptoms often persist even after erroneous beliefs are addressed. Several theorists have proposed alternative ways to understand anxiety-related beliefs, challenging traditional models like CBT. Kelly suggested that beliefs are not static but have qualities such as being tight or loose, affecting how they influence perception. Emotion-focused therapies, on the other hand, argue that emotions precede belief formation, and thus anxiety should be understood through the factors that determine emotions.

Within free energy principle and active inference framewoks anxiety can be re-conceptualized within the framework of working generative models. This view diverges from existing models by focusing on how individuals without initial anxiety may develop anxiety through the process of learning and updating beliefs about uncertainty. The proposed approach aims to unify biological, behaviorist, and cognitivist theories by explaining anxiety as a result of bottom-up belief propagation within a systemic optimization framework. This means that rather than seeing anxiety as stemming from fixed erroneous beliefs, it is viewed as a dynamic process where beliefs about uncertainty are continuously updated based on experiences. This model offers a comprehensive understanding of anxiety formation, bridging gaps between various theoretical perspectives and providing a holistic approach to addressing anxiety disorders.

The free energy principle offers a comprehensive framework for understanding anxiety by conceptualizing it as lack of certainty about the effectiveness of actions in reducing uncertainty levels. According to this principle, anxiety arises when there is a persistent mismatch between expected outcomes of actions and the actual outcomes encountered, leading to an irreducible uncertainty about action policies. This sustained uncertainty impairs the system’s ability to develop adaptive models for minimizing expected free energy, resulting in a perception-action cycle dominated by unpredictable outcomes. Consequently, the system learns to expect uncertainty, which in turn reinforces the perception of an unpredictable world. This learned uncertainty is particularly detrimental because it prevents the system from developing adaptive responses, similar to the concept of learned helplessness. Anxiety can be seen as learned uncertainty regarding the sufficiency of the system’s world model and action policies in achieving preferred outcomes. This uncertainty is derived from feedback on the efficacy of action policies in reducing uncertainty. Even though the system continues to attempt minimizing uncertainty, it lacks adaptive methods to do so, as it perceives the world as inherently unpredictable. This perception removes the epistemic affordance of novelty, which is crucial for resolving uncertainty through expected information gain.

Persistent anxious states, characteristic of anxiety disorders, emerge from the system learning irreducible uncertainty through continuous model updating and action policies. This aligns with experiential accounts of anxiety, where individuals with anxiety disorders perceive the world as unsafe and unpredictable, in contrast to those with lower anxiety levels who view the world as safe and trustworthy. A biological system with a history of persistent uncertainty is likely to expect ambiguous feedback about the effectiveness of its world model and action policies in reducing future uncertainty. This expectation persists regardless of whether the actual uncertainty is reducible, leading to a disproportionate perception of uncertainty. Thus, the system learns that its model and action policies are insufficient for minimizing uncertainty, resulting in the psychological experience of generalized anxiety.

From an affective or emotional inference perspective, anxiety can be understood as part of the generative model that reflects the system’s recognition of its state of uncertainty, manifesting as the thought, “I must be anxious because I cannot decide what to do.” Therefore, experienced anxiety indicates that the biological system has sufficiently expressive generative models to acknowledge its uncertainty. This approach unifies the understanding of anxiety by explaining how continuous exposure to uncertain outcomes leads to a persistent state of learned uncertainty, ultimately resulting in anxiety disorders. Anxiety emerges when there is a persistent mismatch between predicted and actual outcomes, leading to a state of irreducible uncertainty about the effectiveness of one’s actions. In hierarchical modeling, lower-level models process immediate observations while higher-level models integrate these observations to form more stable beliefs over time. For instance, if an agent continuously experiences unpredictable outcomes they receive feedback that their current models are inadequate for reliably predicting outcomes. This leads to a situation where the agent becomes uncertain about the utility of their action policies. This learned uncertainty disrupts the agent’s ability to develop adaptive responses, resulting in a cycle where the agent continually perceives the world as unpredictable, which is a foundational aspect of anxiety disorders.

This process can be understood through Bayesian learning, where the brain updates its beliefs based on new information. When an agent’s experiences consistently indicate that their actions do not lead to predictable outcomes, the agent’s higher-level models adapt to this perceived unpredictability. This creates a belief system where uncertainty becomes a central feature, even if the actual world is not as unpredictable as the agent perceives it to be. Consequently, the agent develops anxious beliefs that persistently expect uncertainty, which interferes with their ability to respond adaptively to new situations. Research shows that early experiences of unpredictability can lead to more severe anxiety symptoms later in life. This suggests that initial exposure to uncertainty can shape how an individual’s brain models the world, leading to a long-term expectation of unpredictability. These early experiences set the stage for the development of generalized anxiety, where the individual continually perceives and reacts to the world as if it were inherently uncertain. In some cases, the world genuinely is uncertain, and modeling this accurately does not necessarily lead to anxiety. However, anxiety arises when the brain inaccurately models the world as uncertain even when it is not. This misalignment occurs when the brain’s models persistently expect uncertainty due to prior experiences, leading to a new state of homeostasis based on learned uncertainty. This pathological state causes the individual to behave as if the world is uncertain, even when it becomes more predictable.

Depression

Much like anxiety, the FEP has been proposed as a framework where depression emerges from a collection of depressive beliefs or distorted representations of the world, particularly related to reward representation. Rumination involves repetitive and often unproductive thoughts about one’s mood, problems, and their origins, which typically leads to a depressed mood and damaged problem-solving. The active inference framework helps explain how rumination, a key aspect of MDD and a marker for comorbidity between depression and anxiety, can lead to pathologies by disrupting the brain’s ability to minimize uncertainty. Rumination’s process is explained through the brain’s use of hierarchical generative models. Mental simulation, or “running imaginary actions,” plays a crucial role in this process, enabling the brain to optimize generative models without overt action. This involves evaluating policies (sequences of actions) based on their expected ability to reduce uncertainty and make accurate predictions.

The process of mental simulation is divided into three steps: generating candidates for simulation from possible experiences, choosing a candidate to focus on, and utilizing the chosen candidate to reduce prediction errors through mental problem-solving. Distortions in any of these steps can result in depressive rumination. Active inference suggests that the brain constructs internal models to predict parameters essential for survival, creating “self-fulfilling prophecies” through its actions. In the context of rumination, distortions in mental simulation can lead to excessive and maladaptive thinking patterns. In other words, the brain fails to update its models adaptively, leading to persistent maladaptive models. This framework emphasizes the importance of understanding rumination not just as a maladaptive cognitive process but as a disruption in the brain’s ability to optimize its generative models, this perspective provides a comprehensive explanation for how rumination develops and persists, leading to depression and other mental disorders.

Several factors contribute to this unsuccessful mental problem-solving, resulting in stabilized rumination. First, the high computational complexity of real-world problems, especially interpersonal and social situations, makes them difficult to solve. Depressed individuals may struggle to approximate solutions due to the high computational demands required for such complex tasks. Additionally, the complex nature of problems involved in rumination often demands high precision, making perfect solutions unattainable and prolonging rumination. Traumatic events, low mood, and anxiety can exacerbate this demand for precision.

Likewise, depressed individuals might overestimate losses and prematurely discard potentially successful paths due to initial minor losses. This results in fewer options being pursued. Furthermore, missing action policies due to limited experiences increase computational efforts and restrict the number of viable solutions. Learned helplessness occurs when repeated failures lead individuals to conclude that uncertainty and/or loss cannot be resolved, reducing their engagement in problem-solving. This shifts their belief system, making them less likely to pursue deeper mental exploration. The lack of alternative behavioral policies due to the generative model’s poor fit reduces the estimated pragmatic value of actions, making individuals less likely to engage in new behaviors. This imbalance leads to more candidate sampling rather than actual problem-solving. As, the process of sampling policy candidates can also be biased and intercorrelated, yielding limited and repetitive policy options. Depressed individuals often remain in similar physical and emotional states, leading to repetitive sampling of similar policy candidates. This feedback loop reduces the likelihood of finding effective solutions. Additionally, poor generative models and inefficient action policies result in less informative sensory data, exacerbating rumination by relying more on cognitively generated prediction errors.

Long-term rumination has several negative consequences such opportunity costs it brings about, as neither pragmatic nor epistemic value is gained during depressive rumination. This leads to a decline in behavioral output and generative models, creating a vicious cycle where biases intensify over time. Brain’s continuous optimization of its knowledge structure is disrupted by making ruminative states a default response. This latter, affects problem-solving and reducing intrinsic motivation. Rumination also distorts both experiential and narrative self-perceptions, making it difficult to update with new information and perpetuating negative self-perceptions. This distorted narrative of self, influences hyperpriors over candidates for thought, further entrenching ruminative patterns.

As a treatment framework along these lines to avoid repetitive sampling, therapists should consider the context, mood, or bodily states during thought and experience sampling. Along the same lines, metacognitive and acceptance and commitment therapies have shown promise by guiding cognitive processes and goal setting based on personal values. Explicitly incorporating problem complexity estimations and analyzing the costs and benefits of mental engagement can further refine treatments. Patients should learn to predict the time and benefits of solving problems through mental simulation compared to overt actions. It is also important to highlight the shared etiology between MDD and anxiety disorders, with rumination being a common factor. Thus, the proposed improvements for treating rumination in MDD could be beneficial for other conditions, such as social anxiety disorder and generalized anxiety disorder. Preventing excessive rumination should be a key therapeutic target across various psychological disorders.

Conclusion

Here we explored the free energy principle and active inference framework as novel approaches to understanding anxiety and depression, with both disorders being conceptualized as issues with belief updating, where maladaptive priors (learned uncertainties), impair the brain’s capacity to minimize uncertainty and predict outcomes effectively. In anxiety, persistent uncertainty regarding the effectiveness of one’s actions leads to an overestimation of unpredictability, even when the world may not be as threatening as perceived. Depression, similarly, involves maladaptive belief systems that reinforce rumination, creating a cycle where individuals fail to adaptively problem-solve, exacerbating depressive symptoms. Despite offering a unifying theoretical framework, the application of FEP to psychiatric conditions has challenges such as the complexity of translating abstract computational models into clinical practice. It is difficult to create therapeutic interventions that directly address maladaptive priors, particularly since these are deeply ingrained and manifest at multiple levels of cognitive and sensory processing. Additionally, there is a need for empirical validation of how the free energy principle applies to individual cases of anxiety and depression, given the heterogeneity of these disorders. Future research should focus on several areas. First, exploring how the brain’s hierarchical generative models can be more precisely measured in clinical populations could help in developing targeted therapies. Second, investigating how different therapies—such as cognitive-behavioral, metacognitive, or analytic therapy can be improved by focusing on belief updating and sensory attenuation could lead to more effective treatments. Lastly, to study the shared mechanisms underlying anxiety and depression, particularly regarding rumination, as this could help create interventions that address multiple disorders simultaneously (i.e. transdaignostic intervention).

References