ICU Delirium as a Failure of Predictive Synchronization: A Two-Agent Active Inference Model

This paper presents a computational model of delirium in the Intensive Care Unit (ICU), in which delirium is defined as the endpoint of a self-reinforcing cycle of predictive failure between two bidirectionally coupled agents: the patient and the ICU room environment. Drawing on the active inference framework and the free energy principle, the paper proposes that delirium is not a property of the patient in isolation but a relational phenomenon that emerges when the environment persistently fails to predict the patient’s internal state. This failure triggers a causal feedback mechanism in which desynchronization pressure progressively sharpens the patient’s prior beliefs—implementing precision rigidity in the correct active inference sense: not a brain overwhelmed by noise but a brain locked into a state that incoming observations can no longer update. The model is implemented as a two-agent POMDP in which both agents maintain generative models and continuously attempt to predict each other’s states. The room agent (R)—understood as the environment-side sensing–inference–actuation loop, whether instantiated by clinical staff or by an automated monitoring system—infers the patient (P)’s latent parameters (Formula presented.) over time and builds a progressively personalized generative model of the patient. Synchronization is operationalized via two commensurable directional surprisal metrics: (Formula presented.), the room’s surprisal at the patient’s true state, and (Formula presented.), the patient’s surprisal at the room’s observations. A systematic ablation study across four model variants shows that room inference is the architectural component necessary to reproduce the synchronization–delirium relationship: when the room infers, the association between synchronization and declared delirium is strong and stable, whereas a non-inferring room collapses to ceiling delirium rates and a weak association. (Formula presented.) learning and the prior-sharpening feedback do not increase the strength of this association; instead they shape the phenotypic gradient, reducing ceiling effects in vulnerable phenotypes and amplifying the separation between them. The model is presented as a computational hypothesis generator rather than a calibrated clinical predictor, and its implications for ICU design are discussed.