Psychiatric disorders as failures in the prediction machine

Due to drastic improvements in computer power and the refinement of machine learning (ML) theories and artificial intelligence (AI) technologies, theoretical and mathematical methods are expected to contribute to psychiatry and clinical neuroscience. This emerging field of research is referred to as ‘computational psychiatry.’ ¹ There are two major strategies for computational psychiatry. The first is trying to discover covert regularity from biological, neurological, and behavioral ‘big data’ related to psychiatric disorders using ML/AI techniques, which is referred to as the data‐driven approach. Although big data and cutting‐edge ML/AI techniques are used, the data‐driven approach is still an extension of conventional methods, in the sense that it explores correspondences (correlations) between observed data and phenotypes and does not examine information processing within the brain itself. Therefore, the data‐driven approach alone may be insufficient to overcome the fundamental difficulties in investigating mental illness, such as biological nonspecificity and heterogeneity. In order to address this issue, there is another line of approach in computational psychiatry, referred to as a theory‐driven approach, in which mental disorders are modeled as aberrant information processing (‘computation’) in the brain using mathematical formulations. The theory‐driven approach is expected to provide mechanistic explanations bridging the different levels of biological observations, including genes, molecules, cells, neural circuits, physiology, behaviors, and symptoms.

In this issue, Smith et al. ² provide a review of the recent advances in the application of the theory‐driven approach in psychiatry and clinical neuroscience research. Specifically, they focus on ‘predictive processing theory’ (also referred to as predictive coding and active inference), which is one of the most promising theories used in the theory‐driven approach. According to predictive processing theory, the brain is a ‘prediction machine’ based on an internal model of the world and interacts with the world through a computational rule of ‘prediction‐error (PE) minimization.’ PE minimization can be achieved in the following three ways: modification of the internal model (learning), modification of internal/brain states (perception/recognition), and modification of sensory inputs (change the world through action). As such, predictive processing theory can explain a wide range of brain functions, including learning, perception/cognition, and action based on PE minimization. While the hierarchical predictive process provides significant advantages for adaptive behavior in social environments, the failure to properly develop or maintain precisely aligned signaling in neural systems has been postulated to result in psychiatric or developmental disorder symptoms. Indeed, regarding failures in predictive processing, several computational modeling studies have provided mechanistic understanding of the perceptual and cognitive impairments in autism spectrum disorders and schizophrenia. ³

Smith et al. ² provide a comprehensive review of recent advances in predictive processing theory, in which, by introducing the concept of ‘interoceptive prediction,’ predictive processing theory has been extended to interoceptive systems. Thanks to this extension of the theory, affective and somatic symptom disorders have fallen within the scope of predictive processing theory. For example, in one of the studies introduced in Smith et al., ² Stephan et al. ⁴ propose a mathematical model of homeostatic and allostatic regulation of visceral states as a hierarchical interoceptive predictive process in which homeostatic adjustment of physiological variables, such as blood osmolality and circulating hormones, can be achieved by a top‐down interoceptive prediction (allostasis) and PE (deviations from the homeostatic range) minimization process. Using this model, they hypothesized that when attempts at higher‐level allostasis repeatedly fail, subjective fatigue and depressive behavior could emerge as strategies for dealing with conditions in which the brain predicts that allostatic regulation will be ineffective. In addition, Smith et al. ² emphasize the importance of mathematical formulation in the sense that, by using quantitative prediction, the model makes it possible to investigate which regions of the brain show patterns of activity consistent with those simulated dynamics of prediction and PE, and identify the differences in individuals, including diagnosis and prognostic information.

Despite increasing expectations, there have not yet been computational modeling studies providing specific effects on clinical practice. For further advances in the application of formal models of predictive processing, the involvement of experts in psychiatry and clinical neuroscience with sophisticated knowledge of symptomatology and neuroscience (i.e., the readers of Psychiatry and Clinical Neurosciences) is crucial.