Modeling the Action-Perception Loop and its role in Phantom Limb Pain using Active Inference

Preprint, 2026

Mathematical models grounded in the Bayesian inference framework have been used to describe various aspects of pain perception. However, pain is not only passively inferred but actively shaped through interactions with the environment – a dimension that classical Bayesian approaches typically do not capture. Here, we present a model within the active inference framework, which extends Bayesian inference to include action selection, to investigate the role of this action-perception loop in phantom limb pain. Despite being one of the most prevalent and distressing consequences of limb amputation, theories regarding the underlying mechanisms of phantom limb pain remain disputed. Because amputation disrupts both sensory input related to the limb and the ability to perform actions, a model incorporating both sensory and active components of pain may provide new mechanistic insight into this condition. The presented model provides a conceptual account of how loss of limb control, ambiguity in sensory input pertaining to limb position, residual noxious input, and pre-amputation pain may contribute to the emergence and persistence of phantom limb pain. Furthermore, the model offers insight into the possible mechanisms underlying common interventions and may help account for their variable efficacy across individuals.