Random forest for prediction of contrast-induced nephropathy following coronary angiography

The majority of prediction models for contrast-induced nephropathy (CIN) have moderate performance. Therefore, we aimed to develop a better pre-procedural prediction tool for CIN following contemporary percutaneous coronary intervention (PCI) or coronary angiography (CAG). A total of 3469 patients undergoing PCI/CAG between January 2010 and December 2013 were randomly divided into a training (n = 2428, 70%) and validation data-sets (n = 1041, 30%). Random forest full models were developed using 40 pre-procedural variables, of which 13 variables were selected for a reduced CIN model. CIN developed in 78 (3.21%) and 37 of patients (3.54%) in the training and validation datasets, respectively. In the validation dataset, the full and reduced models demonstrated improved discrimination over classic Mehran, ACEF CIN risk scores (AUC 0.842 and 0.825 over 0.762 and 0.701, respectively, all P < 0.05) and common estimated glomerular filtration rate. Compared to that for the Mehran risk score model, the full and reduced models had significantly improved fit based on the net reclassification improvement (all P < 0.001) and integrated discrimination improvement (P = 0.001, 0.028, respectively). Using the above models, 2462 (66.7%), 661, and 346 patients were categorized into low (< 1%), moderate (1% to 7%), and high (> 7%) risk groups, respectively. Our pre-procedural CIN risk prediction algorithm (http://cincalc.com) demonstrated good discriminative ability and was well calibrated when validated. Two-thirds of the patients were at low CIN risk, probably needing less peri-procedural preventive strategy; however, the discriminative ability of CIN risk requires further external validation. TRIAL REGISTRATION: ClinicalTrials.gov NCT01400295.