<![CDATA[Clinical deterioration among hospitalized patients remains a major challenge in inpatient care management. Conventional deterioration monitoring approaches frequently rely on isolated physiological observations and static threshold-based scoring systems, which may inadequately represent the longitudinal nature of patient instability. This study aimed to develop and evaluate a retrospective temporal patient risk timeline framework for clinical deterioration surveillance using hospital data warehouse and machine learning. A retrospective longitudinal deterioration surveillance framework was developed using hourly inpatient observations derived from hospital operational and clinical data sources. A temporal XGBoost model was developed to predict ICU transfer within the subsequent 24-hour observation window under severe class imbalance conditions. To evaluate the additional value of temporal feature engineering, baseline non-temporal models including logistic regression, random forest, and non-temporal XGBoost were also developed using static observation-level variables. Operational feasibility was retrospectively demonstrated through a dashboard prototype integrating deterioration prioritization, patient timeline visualization, and longitudinal surveillance. The final dataset consisted of 20,181 inpatient admission episodes represented by 2,233,143 hourly observations. ICU transfer occurred in 712 admissions. Because the prediction task was formulated as hourly deterioration surveillance, ICU transfer labels represented only 0.65% of all hourly observations, indicating severe class imbalance. The temporal XGBoost model achieved an AUROC of 0.606 and an AUPRC of 0.009. While predictive discrimination was modest, the framework enabled continuous longitudinal representation of patient deterioration and retrospective surveillance of clinical instability. The findings suggest that temporal patient representation may provide operational and interpretative value beyond static snapshot-based deterioration monitoring approaches.]]>
