Belief-State Design for ICU Deterioration Forecasting Under Endogenous Observation and Scarce Review Capacity

• Nguyen Minh Khoa

  Department of Computer Engineering, Hung Yen University of Technology and Education, Khoai Chau Campus, Dan Tien Road, Dan Tien Commune, Khoai Chau District, Hung Yen, Vietnam

  Author

• Tran Quoc Huy

  Faculty of Information Technology, Ha Tinh University, 447 26/3 Street, Dai Nai Ward, Ha Tinh City, Ha Tinh, Vietnam

  Author

Critical care forecasting is often presented as a straightforward supervised learning task in which future deterioration is predicted from current electronic health record features. That framing is serviceable for benchmarking, but it hides a defining computational property of intensive care data: what becomes visible in the record is shaped by clinical attention itself. Measurements are ordered selectively, documentation density changes with concern and workflow, interventions alter both patient state and what is subsequently monitored, and the absence of information is frequently informative about the observation process rather than the patient alone. As a result, the central technical challenge is not simply prediction from incomplete data. It is inference over a partially observed dynamical system whose observation policy is endogenous to the underlying state and to clinician action. This paper develops a computer-science-centered framework that redefines ICU deterioration modeling as belief-state design under endogenous observation. The proposed view separates latent physiologic state, observation policy, intervention context, confidence, and marginal review value. A patient representation is then learned not merely to maximize event discrimination, but to support risk estimation, uncertainty formation, and prioritized review under constrained clinical capacity. The paper introduces a state-space formulation in which structured measurements, missingness patterns, and clinical notes are interpreted jointly through an observation model, a belief update mechanism, and a queue-facing utility layer. It also examines how rare-event training, temporal dependence, calibration, and support drift interact with the allocation of scarce human attention. The resulting argument is that useful ICU forecasting systems should be designed less as binary event detectors and more as engines for maintaining and acting on belief states when both the patient and the record are evolving under policy.

• PDF