Abstract Rationale Early identification of hospitalized patients at risk of clinical deterioration is essential for timely intervention and optimal resource allocation. Traditional early warning systems, which rely on limited physiological parameters and static thresholds, often fail to capture the complex and dynamic patterns that precede intensive care needs. This study aimed to develop and validate an artificial intelligence (AI) model using routinely collected electronic medical record (EMR) data to predict the likelihood of critical deterioration and intensive care unit (ICU) transfer among inpatients. Methods We retrospectively included adult inpatients admitted to a tertiary medical center between January 2021 and December 2024. EMR data, including demographics, vital signs, and laboratory results, were extracted as model features. The primary outcome was unplanned in-hospital deterioration requiring ICU transfer within 24 hours. Several machine learning algorithms, including Light Gradient Boosting Machine (LGBM) and eXtreme Gradient Boosting (XGBoost), were trained and evaluated. Model performance was assessed by the area under the receiver operating characteristic curve (AUROC), precision, recall, and F1 score. Model interpretability was examined using Shapley Additive Explanations (SHAP) analysis. Results A total of 10,879 patients were included, contributing 82,404 records. The final XGBoost model achieved the best predictive performance, with an AUROC of 0.8708, precision of 0.7907, recall of 0.8068, and F1 score of 0.7987. The most influential predictors included blood pressure, respiratory rate, heary rate, use of mechanical ventilation, blood neutrophil count, blood eosinophil count, C-reactive protein level, serum aspartate aminotransferase level, serum alanine aminotransferase level, serum lactate level, serum creatinine level, and so on. Conclusions We developed an EMR-based AI model that accurately predicts inpatient clinical deterioration and potential ICU needs. This model provides a real-time, data-driven decision support tool for hospital care teams. Integration of such predictive systems into clinical workflows may enhance situational awareness, facilitate timely intervention, optimize ICU resource utilization, and ultimately improve patient outcomes. This abstract is funded by: Kaohsiung Medical University Hospital (SH11104, SH11204, SH11301, SH11401)
Tsai et al. (Fri,) studied this question.