Contrast-Induced Acute Kidney Injury in Lower Limb Percutaneous Transluminal Angioplasty: A Machine Learning Approach for Preoperative Risk Prediction.

BACKGROUND: Contrast-induced acute kidney injury (CI-AKI) is a common complication of lower limb percutaneous transluminal angioplasty (PTA). Common risk models are based on cardiology cohorts for percutaneous coronary intervention. They include a mix of preoperative and perioperative variables, but do not include important information such as inflammatory parameters and preoperative medications. None make use of machine learning. We aimed to develop an accurate preoperative risk model for CI-AKI in lower limb PTA using machine learning methods and comparing these with conventional logistic regression.

MATERIALS AND METHODS: A retrospective cohort of 456 patients who underwent lower limb PTA as an isolated procedure from 2015 to 2019 was identified. Patients <21 years old, patients with a preoperative estimated glomerular filtration rate of <15 mL/min/1.73 m2 as defined by the modification of diet in renal disease, and patients with no valid preoperative or postoperative serum creatinine were excluded. Conventional logistic regression and a range of machine learning models were fitted (logistic regression with elastic-net penalty, random forests, gradient boosting machines, k-nearest neighbors, Support vector machines, and multilayer perceptron), using 5-fold cross-validation and grid search for hyperparameter selection. Area under receiver operating curve, area under precision-recall curve, F1 score, and the sensitivity and specificity were determined on the test set. Variable importance was examined using SHapley Additive exPlanation plots.

RESULTS: Machine learning models performed well, with the best performance by the k-nearest neighbors algorithm (area under receiver operating curve = 0.914, area under precision-recall curve = 0.809). Important variables identified by SHapley Additive exPlanation plot analysis included modification of diet in renal disease estimated glomerular filtration rate, haemoglobin, and inflammatory indices (neutrophil: lymphocyte ratio, red cell distribution width).

CONCLUSION: We developed machine learning models to accurately predict CI-AKI in patients undergoing elective lower limb PTA, using preoperative variables only. This model may be used for preoperative patient risk counseling by surgeons and anesthetists and may assist in identifying high-risk patients for further monitoring.