Machine Learning Algorithms for Predicting and Risk Profiling of Cardiac Surgery-Associated Acute Kidney Injury

Abstract

Using a large national database of cardiac surgical procedures, we applied machine learning (ML) to risk stratification and profiling for cardiac surgery-associated acute kidney injury. We compared performance of ML to established scoring tools. Four ML algorithms were used, including logistic regression (LR), gradient boosted machine (GBM), K-nearest neighbor, and neural networks (NN). These were compared to the Cleveland Clinic score, and a risk score developed on the same database. Five-fold cross-validation repeated 20 times was used to measure the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. Risk profiles from GBM and NN were generated using Shapley additive values. A total of 97,964 surgery events in 96,653 patients were included. For predicting postoperative renal replacement therapy using pre- and intraoperative data, LR, GBM, and NN achieved an AUC (standard deviation) of 0.84 (0.01), 0.85 (0.01), 0.84 (0.01) respectively outperforming the highest performing scoring tool with 0.81 (0.004). For predicting cardiac surgery-associated acute kidney injury, LR, GBM, and NN each achieved 0.77 (0.01), 0.78 (0.01), 0.77 (0.01) respectively outperforming the scoring tool with 0.75 (0.004). Compared to scores and LR, shapely additive values analysis of black box model predictions was able to generate patient-level explanations for each prediction. ML algorithms provide state-of-the-art approaches to risk stratification. Explanatory modeling can exploit complex decision boundaries to aid the clinician in understanding the risks specific to individual patients.