Modeling and controlling the body in maladaptive ways: an active inference perspective on non-suicidal self-injury behaviors

Abstract

A significant number of persons engage in paradoxical behaviors, such as extreme food restriction (up to starvation) and non-suicidal self-injuries, especially during periods of rapid changes, such as adolescence. Here, we contextualize these and related paradoxical behavior within an active inference view of brain functions, which assumes that the brain forms predictive models of bodily variables, emotional experiences, and the embodied self and continuously strives to reduce the uncertainty of such models. We propose that not only in conditions of excessive or prolonged uncertainty, such as in clinical conditions, but also during pivotal periods of developmental transition, paradoxical behaviors might emerge as maladaptive strategies to reduce uncertainty-by "acting on the body"- soliciting salient perceptual and interoceptive sensations, such as pain or excessive levels of hunger. Although such strategies are maladaptive and run against our basic homeostatic imperatives, they might be functional not only to provide some short-term reward (e.g. relief from emotional distress)-as previously proposed-but also to reduce uncertainty and possibly to restore a coherent model of one's bodily experience and the self, affording greater confidence in who we are and what course of actions we should pursue.

Keywords: active inference; adolescence; interoception; intolerance of uncertainty; non-suicidal self-injuries.

Figure 1.

A schematic illustration of a hierarchical active inference model. This model links (exteroceptive, interoceptive, and proprioceptive) sensations at lower levels with multimodal models of hidden bodily states, such as fatigue and hunger, at intermediate levels, and finally with temporally extended, integrative models of the embodied self at the higher hierarchical level. In this schematic, following predictive coding (Rao and Ballard 1999, Friston 2005), black and red circles represent neural units that encode predictions and prediction errors, respectively. The levels are reciprocally connected, so predictions are propagated from the top-down (black edges) and prediction errors from the bottom-up (red edges). Finally, the pink triangles indicate a mechanism of precision gating (or gain control) of prediction error units, which determines their relative influence on units encoding predictions. At a neurobiological level, prediction and prediction error units could be mapped to deep and superficial pyramidal cells in cortical hierarchies, whereas expected precision could be linked to neuromodulatory input. The elements of the generative model shown do not need to map one-to-one to specific brain areas or networks but are plausibly distributed across many of them. However, as a first approximation, the lower and intermediate layers of the generative model could be linked to brain networks that process unimodal information (e.g. sensory cortices for exteroceptive information) and multimodal association areas, respectively. The highest level of the generative model could be linked to brain networks that process information about the self, such as the insular cortex, the anterior cingulate cortex, and the medial prefrontal cortex. See Parr et al. (2022) for details about hierarchical generative models supporting adaptive regulation and allostasis and Barrett and Simmons (2015) for their putative neuronal underpinnings. See online article for colored version of this figure.

Figure 2.

A simplified example of (Bayesian) inference of one’s heart rate. First panel: simulated time series of heartbeat observations. Second panel: Shannon surprise of a generative model composed of a fixed prior about heart rate (a Gaussian with a mean of 67 and a precision of 0.11) and a likelihood (a Gaussian centered on the current heart rate with an additional bias of 15 pulses, with various precisions that vary between 0.47 and 10, see the legend). Third panel: Bayesian surprise, which measures the discrepancy between posterior and prior probabilities over time. Bottom panels: the two series of panels are organized in two (left and right) columns, which show the first five time steps of inference for the two cases with high precision (of 10) and low precision (of 0.1) of the likelihood, respectively. See the main text for an explanation and online article for colored version of this figure.