Case Report: Anti-N-methyl-D-aspartate receptor encephalitis associated with bilateral mature ovarian teratomas in early pregnancy.

Predictive processing theories propose that the brain does not passively register sensory information but actively anticipates and interprets incoming stimuli through the continuous generation, monitoring, and updating of internal neural models. Two central components of these theories are predictions and prediction errors: predictions correspond to the brain’s expectations about forthcoming sensory input, whereas prediction error signals represent the discrepancies between expected and actual sensory events (; ). Within this framework, higher cortical areas generate top-down predictions that are conveyed to lower hierarchical levels in order to suppress neuronal responses to expected events. When incoming sensory input deviates from current predictions, lower levels transmit bottom-up prediction errors to higher hierarchical levels, enabling the system to update its model of the causes of sensory input (; ; ; ). Through this recurrent message passing, prediction errors are progressively minimized across the hierarchy (). Hierarchical predictive coding scheme. Differences between the expected sensory input and real input generate ascending prediction errors (red arrows) that are transmitted through cortical hierarchies to update higher-level expectations. These expectations generate descending predictions that suppress predictable sensory input (blue arrows). Prediction errors are weighted through the neuromodulatory systems (green) according to their expected precision, such that highly precise signals exert greater influence on perceptual updating, whereas imprecise signals are attenuated. Through this precision-dependent gain control, the system dynamically balances bottom-up sensory evidence and top-down predictions, selectively enhancing informative signals while suppressing unreliable or noisy information. This process contributes to “representational sharpening,” whereby relevant sensory representations become more sensitive to incoming input. This framework conceptualizes the brain as a statistical inference machine (; ). In contemporary free-energy formulations, this idea is implemented as a biologically plausible account of perceptual inference, in which the brain embodies a generative model of the environment that captures the hierarchical and dynamic structure of sensory input. Neuronal activity is therefore understood as reflecting inference about the hidden causes of sensory events, allowing the brain to predict its own sensations (; ; ). In Bayesian terms, optimal recognition requires the brain to represent probability distributions over hidden states and environmental causes, updating these representations until they approximate the statistical regularities of the world. Crucially, such inference depends not only on the content of predictions, but also on the precision assigned to prediction errors at different levels of the hierarchy.