A Decision Architecture for Safety Computations

Abstract

Accurately estimating safety is critical to pursuing nondefensive survival behaviors. However, little attention has been paid to how the human brain computes safety. We conceptualize a model that consists of two components: (i) threat-oriented evaluations that focus on threat value, imminence, and predictability; and (ii) self-oriented evaluations that focus on the agent's experience, strategies, and ability to control the situation. Our model points to the dynamic interaction between these two components as a mechanism of safety estimation. Based on a growing body of human literature, we hypothesize that distinct regions of the ventromedial prefrontal cortex (vmPFC) respond to threat and safety to facilitate survival decisions. We suggest safety is not an inverse of danger, but reflects independent computations that mediate defensive circuits and behaviors.

Keywords: decision making; safety; threat; ventromedial prefrontal cortex.

Figure 1.. Protection and imminence influence safety estimates.

A. Threat imminence continuum (TIC). Based on Fanselow and Lester (1988). B. Protection alters safety perceptions along the TIC even as attack probability remains constant. C. Behavioral and emotional responses to threat imminence vary as a function of safety perception. Emotion terms for low safety contexts were based on Mobbs et al. (2020) and positively valenced emotions with similar arousal at the same level were selected for high safety contexts using the Dictionary of Affect in Language (DAL; [97]). MB=model based, MF=model free.

Figure 2,. Key Figure. Safety Decision Model components and computational flexibility across the safety continuum.

A. A hypothetical rendering of the Safety Decision Model components underlying safety computation. Solid black arrows indicate temporal influence. Evaluative components exert bidirectional influence on other lower-level features. Safety computation: Relative titrating of threat- and self-oriented evaluative components are integrated in estimates of survival to compute perceived safety. Safety signal: Distinct predictions about survival success are encoded at the neural level. Safety behavior: As safety increases, non-defensive survival behaviors are prioritized. As safety decreases, threat monitoring and defensive behaviors increase, suppressing non-defensive motivated behavior. Safety outcome: Prediction errors are integrated to update future safety computations. B. Safety computation factors are proposed to vary in the extent to which they are processed along the safety continuum.

Figure 3.. Safety Decision Neural Model and evidence supporting a posterior-to-anterior threat-to-safety distinction.

A. Simplified model of the proposed flow of threat- and self-oriented processes from the Safety Decision Model resulting in safety computation. Green illustrates anterior vmPFC safety signal. Regions in yellow (underlying self-oriented evaluations) and red (underlying threat-oriented evaluations) alter responding depending on safety signaling. In low safety contexts these regions are involved in defensive responding whereas in safe contexts they promote non-defensive behavior. Grey regions are part of the canonical defensive circuit that are suppressed when safety is computed. Arrows represent the flow of functional safety pathways, but do not suggest comprehensive functional or anatomical connectivity pathways. B. Peak coordinates from representative human functional magnetic resonance imaging studies reporting neural response to safety (green) and threat (red) (see Supplemental Table S1 for study details). Posterior-vmPFC (red) and anterior vmPFC (green) seeds are overlayed as circles. C. Meta-analytic decoding with Neurosynth. Red and green radar bars represent correlation strength between key words representing components of the Safety Decision Model and the anterior (green) and posterior vmPFC (red) ROIs. D. Meta-analytic coactivations from Neurosynth including 14371 studies for ROIs in the anterior (green; x=−2, y=46, z=−10) and posterior (red; x=0, y=26, z=−12) vmPFC. vmPFC=ventromedial prefrontal cortex, BLA=basolateral amygdala, CeA=central amygdala, PAG=periaqueductal gray, MCC= midcingulate cortex, hypothal= hypothalamus, ant=anterior, post=posterior.