Deployment of an AI Model for Predicting Acute Kidney Injury in Hospitalized Patients
Abstract
Acute kidney injury (AKI) has a prevalence of about 10-15% among hospitalized patients and is a cause of substantial morbidity, mortality, and healthcare expenditure but also has a limited ability to detect it early since conventional indicators have delayed response times. The present research is an attempt to create and implement a machine learning regimen to predict the onset of AKI 12-48 hours before clinical presentation with the available electronic health record data. The model used was a gradient boosting ensemble based on data of 15,847 patients who were hospitalized and the attributes used included 127 clinical variables such as vital signs, laboratory values, medications, and demographically related factors. Good predictive performance was attained by the model with an area under the receiver operating characteristic curve (AUROC) of 0.892, sensitivity of 84.3% and specificity of 81.7% on the validation set. The cross-validation analysis proved strong generalizability of the model into various populations of patients and diverse units of the hospital. The implementation of this AI model in clinical care will provide the ability to intervene before the actual development of the condition, mitigate the adverse outcomes on AKI complications, and will benefit patient outcomes in terms of decreased healthcare expenditures instead.