Continuous Risk Prioritization for ICU Deterioration Under Limited Clinical Attention

• Bilal Mahmood

  Department of Computer Science, University of Chakwal, Talagang Road, Chakwal 48800, Pakistan

  Author

• Tariq Mehmood

  Department of Software Engineering, University of Swabi, Anbar Road, Swabi 23430, Pakistan

  Author

Critical care units have become increasingly instrumented environments, but the practical value of prediction depends not only on whether deterioration can be anticipated, but on whether scarce clinical attention can be directed toward the right patients at the right time. In most machine learning studies, deterioration is framed as a binary forecasting problem, and model quality is summarized with threshold-free discrimination metrics. That framing is useful for benchmarking, yet it obscures the operational structure of intensive care surveillance. Clinicians do not respond to isolated probabilities in the abstract. They allocate monitoring effort, bedside reassessment, diagnostic work, and intervention readiness across a continuously changing population of patients under hard constraints on time and staffing. This paper develops a computer-science-centered formulation of ICU deterioration modeling as a continuous risk prioritization problem rather than as a simple binary alarm problem. The framework treats each patient as a partially observed stochastic process whose risk estimate determines position in a dynamic ranked worklist. Emphasis is placed on queue-constrained surveillance, temporal smoothness, uncertainty-aware escalation, and the geometry of the high-risk tail where human review capacity is actually spent. A mathematical treatment is introduced for dynamic priority scoring, workload-limited thresholding, and risk propagation under intervention-sensitive trajectories. The paper also examines how rare-event learning, calibration, selective abstention, and distribution shift interact with ranking-based deployment. The central argument is that deterioration systems are best evaluated and designed as streaming prioritization engines embedded in an attention-limited clinical environment, where useful prediction is inseparable from scheduling, reliability, and operational control.

• PDF