摘要：Background: Sepsis is vital in critical care medicine, and early detection and intervention are key to survival. We aimed to establish an early warning system for sepsis based on a data integration system that can be implemented in the intensive care unit (ICU). Methods: We trained the LightGBM and multilayer perceptron on the open-source database Medical Information Mart for Intensive Care for sepsis prediction. An ensemble sepsis prediction model was established based on the transfer learning and ensemble learning technique on the private dataset of Ruijin Hospital. The Shapley Additive Explanations analysis was applied to present feature importance on the prediction inference. With the development of data-integrating hub to collect and transmit data from different brands of ICU medical devices, the data integration system was established to receive, integrate, standardize, and store the real-time clinical data. In this way, the sepsis prediction model developed in the ICU of the Ruijin Hospital for the real-world study of sepsis early warning on ICU management. The trial was registered with ClinicalTrials.gov (NCT05088850). Findings: Our best early warning model achieved an area under the receiver operating characteristic curve (AUC) of 0·9833 in the task of detecting sepsis in 4-h preceding on the open-source database, while our ensemble model achieved an AUC of 0·90650·9436 in the retrospective research from 15-h preceding on the private database, and 0·86360·8992 in real-time real-world studies using the data integration system in the ICU of the Ruijin Hospital. In the continuous early warning process of patients admitted to the ICU, 22 patients who met the diagnostic criteria for sepsis during hospitalization were predicted as positive cases; 29 patients without sepsis were predicted as negative cases. Additionally, 17 patients were predicted as false-positive cases; in six patients with sepsis during ICU stay, the predicted probabilities at different time nodes were all less than the warning threshold 0·7 and predicted as false-negative cases. Interpretation: Machine learning models could allow accurate and real-time inference to detect sepsis onset within 5-h preceding at most with the help of the data integration system. We identified the features such as age, antibiotics, ventilation, and net balance to be important for the sepsis prediction inference. We argue that this system has promising potential to improve ICU management by helping medical practitioners identify at-sepsis-risk patients and prepare for timely diagnosis and intervention. Funding: Shanghai Municipal Science and Technology Major Project, the ZHANGJIANG LAB, and the Science and Technology Commission of Shanghai Municipality.