Acute kidney injury (AKI) is a significant issue in ICU care. To improve the quality of care, we developed an AI model to predict AKI risk using clinical data. The steps are as follows: (1) We trained an AI model on data from 13,888 ICU admissions (2015-2019) and validated it with 2,897 records from 2020. The model utilized 60 features, including clinical physiological, medication, and laboratory data. The prediction performance reached AUROC:0.921(2015-2019), 0.928(2020). (2) Using the Lasso method, we selected 21 significant features for model training. The overall AKI prediction can still reach AUROC:0.911(2015-2019). (3) We cooperated with four medical centers by providing model parameters for external validation. The overall AUROC reached at least 0.833. (4) We developed a new model using a federated learning platform with five hospitals through Advantech technology. All hospitals made predictions on their data, with results surpassing single-model external validation. (5) The AKI model was deployed on the WISE-PaaS platform. For risk prob
ability inference, real-time vital signs, medication, and related data were collected upon patient admission. The inferred AKI risk prediction was integrated into the dashboard, offering nephrotoxic drug alerts and drug dose suggestions. (6) A data transfer tool (ETL: Extractor, Transfer & Loader) was employed to establish data transfer jobs. Each job transferred data to the WISE-PaaS Database for inference. The inferred AKI prediction was updated in the hospital's clinical medical information system (EHIS), making it convenient for clinical staff. (7) We integrated the technology and results into a product called "AI Acute Kidney Injury Prediction: Real-time Inference Interactive Critical Care System." This product easily incorporates into hospital information systems for clinical use. The product prototype is complete, and software medical device certification is ongoing.