Using Intraoperative Variables to Predict Acute Kidney Injury Following Cardiac Surgery

After undergoing cardiac surgery, a significant number of patients develop Acute Kidney Injury (AKI), a condition that contributes to higher mortality and morbidity rates. Current methods of diagnosing AKI are largely reactionary, as kidney damage can only be assessed after creatinine levels in the blood rise, a process that occurs 24-48 hours after initial injury. During this time period, doctors make medical decisions that may add extra stress to kidney function, unknowingly contributing to further kidney damage. The University of Virginia (UVa) Health System is interested in improving its ability to predict AKI following cardiac surgery in order to more quickly and accurately identify at-risk patients. Currently, the UVa Health System uses the Society of Thoracic Surgeons (STS) preoperative AKI Risk Score to assess each patient's risk of kidney injury prior to surgery. Hoping to improve predictive performance, the Health System desires a new risk model that also incorporates risk factors from the intraoperative period. The final dataset ($\mathrm{n}=335$ surgeries) includes both preoperative and intraoperative factors compiled from the UVa Health System EMR database. Machine learning models were utilized to predict each patient's change in creatinine level, the metric used to assign AKI classifications. Specific focus was given to incorporating intraoperative time series factors. Changepoint analysis, estimated entropy, and heteroscedastic modeling were employed to analyze the time series readings from lab, anesthesiology, and medication records taken during cardiac surgery. Several of these intraoperative time series features were significant variables in all of the highest performing L1 Linear Regression, L1 Logistic Regression, Random Forest, Neural Net, and Extreme Gradient Boost models.