Abstract No. 482 Machine learning–based prediction of short-term major amputation risk following peripheral arterial endovascular intervention

Abstract

To predict short-term amputation risk using machine learning techniques in patients undergoing lower extremity endovascular interventions for peripheral artery disease In this IRB-approved retrospective study, we utilized the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) general and targeted databases to collect preoperative clinical and lab information on the 2008 patients who underwent lower extremity endovascular procedures for peripheral artery disease between 2014 and 2017. Our primary outcome of interest was 30-day ipsilateral amputation of the affected limb. Features were selected using a recursive random forest feature importance model trained on all available variables in the ACS-NSQIP database. Using a 4:1 training:testing split, we developed and tested multiple machine learning models to predict the occurrence of 30-day amputation following infrainguinal endovascular procedures. All models were tuned using a Bayesian model-based optimization scheme to maximize AU-ROC in 10-fold stratified cross-validation training on the training dataset. Model performance was assessed using calibration plots, AU-ROC (area under the receiver operating curve), AU-PRC (area under the precision-recall curve), F1, Brier score, and Kappa statistics. For the final chosen model, we calculated additional feature importance statistics to offer an explainable model. As a final test, we compared our chosen model to a logistic regression model built on the Rutherford classification scheme present within the ACS-NSQIP database. Using 35 preoperative and demographic characteristics, a machine learning model built using an extreme gradient boosting algorithm achieved the best performance with an AU-ROC of 0.85, a Brier score of 0.04, an F1 score of 0.30, a Kappa of 0.27, an AU-PRC of .26, and an overall accuracy of 0.96. The most important features of the model were white blood cell count, creatinine, partial thromboplastin time, alkaline phosphatase level, and hematocrit. The logistic regression-based Rutherford model achieved an AU-ROC of 0.57. The multifactorial machine learning model we propose performed significantly better than the Rutherford-based model for predicting 30-day amputation risk (DeLong’s P < 0.001). We present a machine learning model that effectively predicts 30-day amputation events in patients undergoing lower extremity endovascular procedures. Such a model can aid in clinical management decisions in patients afflicted with PAD.