Service-Aware Continuous Prioritization for ICU Deterioration Under Queue and Attention Constraints

Abstract

Intensive care prediction is typically framed as binary event forecasting, yet bedside surveillance is rarely practiced as a sequence of isolated yes-or-no decisions. What clinicians actually manage is a continuously changing allocation problem: which patients deserve immediate review, which require intensified monitoring, which are rising in urgency, and which can be safely deprioritized for the moment. This distinction matters because the statistical object most models produce is not the object the unit consumes. The practical output of a deterioration system is a ranked and repeatedly updated worklist constrained by finite review capacity, uncertain evidence quality, intervention saturation, and queue spillover effects. This paper develops a computer-science-centered treatment of ICU deterioration modeling as service-aware continuous prioritization rather than simple binary alarm generation. The framework formalizes the active census as a coupled collection of partially observed stochastic processes whose scores compete for a limited number of review slots. It then derives a dynamic priority model that integrates irregular physiologic streams, time-varying intervention context, note refresh dynamics, and support-aware uncertainty into a population-level queueing formulation. Particular attention is paid to the geometry of the actionable tail, the distinction between risk and review value, and the way calibration, score smoothness, and uncertainty interact with constrained service capacity. The resulting perspective connects sequence modeling, ranking theory, state estimation, queue control, and human-in-the-loop governance. Rather than treating deterioration prediction as a narrow classification task, the paper argues that clinically useful systems should be designed as streaming prioritization engines whose outputs remain stable, interpretable, and capacity-aware under the operational realities of intensive care.