Establishing AKI Warning System in Peking Union Medical College Hospital from a Machine Learning Approach: A Single-center Research Protocol

Abstract

Background and Objective In-hospital acute kidney injury (AKI) has a significant negative impact on patients' outcome and the length of hospital stay. It is significantly important to use the early warning of electronic medical records (EMR) to identify and intervene AKI in a timely manner so as to reduce the severity of AKI and to improve the prognosis of patients. At present, AKI-related research based on the EMR system mainly uses traditional statistical methods for retrospective analysis, mainly for inpatients in single-disciplinary wards, and there is still a lack of early warning models of AKI risk based on artificial intelligence technology in large-scale multi-disciplinary wards with time-sensitive information and further prospective research. This study aims to develop a multiple-ward AKI prediction model tailored for general hospitals in China based on machine-learning algorithms and big data acquired by the EMR system. Methods This single-center study consists of both a retrospective observational study and a prospective study. All hospitalized adult patients admitted in Peking Union Medical College Hospital (PUMCH) between 2016 and 2020 were included in the retrospective study. Logistic regression, naive Bayes, random forest, support vector machine, gradient boosting and recurrent neural network will be used for modeling based on demographics, clinical feature, vital signs, imaging, lab results and hospitalized medical records, which aims to predict AKI 24-48 h in advance and will be internally validated. The prospective study intends to include all adult inpatients in PUMCH for 12 consecutive months. Among them, all adult hospitalized patients within 6 months before the AKI early warning system is launched will be of the control group, and all adult hospitalized patients within 6 months after the AKI early warning system is launched will be of the intervention group. In the intervention group, the AKI early warning system will be embedded in the EMR, and all patients hospitalized for more than 24 hours will be assessed for AKI risk in the next 48 hours in real time every 6 hours. Early intervention will be carried out for high-risk patients. The control group does not have above-mentioned high-risk and alarm prompts of AKI, and no corresponding intervention measures. The incidence of AKI and AKI grade 3, AKI remission rate, end-stage renal disease progression rate, mortality during hospitalization, length of stay, hospitalization expenses and other indicators will be compared between the two groups. Expected Results An estimated number of 127 000 in-hospital patients will be included in the retrospective study, among which 14 605 patients suffer from AKI. The prediction model is expected to predict AKI 24-48 h in advance and the aim for area under receiver operating characteristics curve should be > 0.80. In the prospective study, 34 748 inpatients will be enrolled, including 17 374 in both the intervention group and the control group. The duration time of renal replacement therapy and length of hospital stay in the intervention group should be shorter than those in the control group (P < 0.05); the proportion of renal replacement therapy, the incidence of AKI and AKI 3, the rate of progression of end-stage renal disease, the mortality rate during hospitalization, and the hospitalization cost should be lower than those in the control group (P < 0.05), and the AKI remission rate should be higher than that in the control group (P < 0.05). Expected conclusion EMR-based multi-ward AKI prediction model will predict AKI risk 24-48 h in advance, which will lower AKI incidence and severity, and improve clinical outcomes.

Keywords

• acute kidney injury
• machine learning
• multi-ward
• prediction model