ICCA-Based ICU Physiological State Space Monitor

Critically ill patients can deteriorate rapidly across multiple organ systems. Most intensive care unit (ICU) risk tools rely on measurements collected at a single time point and may not fully capture how physiology evolves or how quickly a patient recovers from disturbance. This single-center retrospective observational study will use routinely collected data from the ICCA reporting database at Zhongshan Hospital, Fudan University to develop and internally validate a research prototype called the ICU Physiological State Space Monitor. Adult ICU admissions will be represented as daily state vectors across 10 physiological domains. The study will characterize each patient's position and movement in a multidimensional state space, identify high-risk regions and possible critical transitions, and quantify physiological resilience using trajectory features such as variability, autocorrelation, curvature, recovery slope, and cross-domain coupling. The primary validation outcome is a composite clinical deterioration event within 72 hours after an eligible index patient-day. The monitor is an analytic and visualization framework for retrospective research and will not be deployed for real-time clinical decision-making in this study.

Inclusion Criteria: * Age 18 years or older. * Admission to an ICU at Zhongshan Hospital, Fudan University between September 2021 and May 31, 2026, with a uniquely identifiable ICU encounter in the ICCA reporting-layer database. * ICU length of stay of at least 24 hours. * Availability of core variables from at least four physiological domains and at least one patient-day suitable for modeling. Exclusion Criteria: * Duplicate import of the same ICU encounter that cannot be reliably deduplicated. * Severe missingness, inconsistency, or irreconcilable error in key identifiers, ICU admission/discharge times, or core timestamps. * Test data, demonstration data, or records determined through data governance to be non-genuine patient records. * Inability to lock the primary source mapping for key variables, or overall data quality insufficient to support construction of the physiological state vector.