Identification of acute kidney injury after non-cardiac surgery by machine learning leveraging a large electronic health record database

Abstract Post-operative acute kidney injury (PO-AKI) is a frequent complication after non-cardiac surgery, substantially increasing patients’ mortality and morbidity rates. Early identification of patients at high risk of PO-AKI could facilitate improved patient care. We investigated a large, single-center database of 13,890 patients undergoing non-cardiac surgery, of whom 1,718 (12.37 $$\%$$ ) developed PO-AKI. Based on 62 pre-, intra- and postoperative patient characteristics, including time-series mean arterial pressure measurements, we trained and validated eight different machine learning (ML) algorithms to predict PO-AKI. Top ML model performance based on Area under the Receiver Operating Characteristics curve (ROC-AUC) was achieved at 0.742 (95% CI, 0.725-0.762) on the validation data set by a Random Forest classifier. Feature Importance assessment revealed a wide range of patient characteristics predictive of PO-AKI, which, however, varied considerably between ensemble and non-ensemble ML models. Training and evaluation of Random Forest models at seven different perioperative time-points revealed a performance increase in ROC-AUC from 0.722 (95% CI, 0.703-0.746) preoperatively to 0.738 (95% CI, 0.716-0.755) postoperatively on the validation data set. In conclusion, we present a comprehensive performance evaluation of different ML algorithms for the personalized prediction of PO-AKI on a large, real-world electronic health record data set.