Machine Learning Model Predicts Open Bypass Outcomes in VQI Claudicants

Abstract

Peripheral artery disease (PAD) studies have largely focused on patients with acute or critical limb ischemia, although one in five patients with PAD are claudicants. New endovascular technology has fallen into favor for use in patients with claudication. However, there is debate surrounding the longevity of these interventions that may ultimately still result in a patient requiring an open bypass that could possible last a patient's lifetime. The aim of this prospective study was to leverage existing data in the Vascular Quality Initiative (VQI) database and preoperatively risk stratify patients with lifestyle-limiting claudication who may be eligible for an open bypass. The VQI was queried for all patients who underwent infrainguinal lower extremity bypass between December 2004 and December 2017. Of the 71,561 procedures, 15,633 of these procedures were initial open procedures performed for patients with lifestyle-limiting claudication. Our outcomes of interest included 30-day survival, major adverse limb events (MALE), and major cardiac events (MCE). The machine learning algorithm Random Forrest was used to identify univariate significant predictors after multiple testing correction. Generalized liner models with all significant predictors were then trained on 80% of the existing data set and tested on 20% of the dataset for accuracy of predicting outcomes. Eleven percent of patients survived after 30 days, 3.2% had an MALE, and 14% had an MCE. The best performing prediction model created was for 30-day survival, where we included the following predictors identified on univariate analysis in addition to age, race, and gender: utilizing a conduit other than great saphenous vein (GSV) (odds ratio [OR], 0.79; 95% confidence interval [CI], 0.72-0.88; P corrected = .05), uncontrolled hypertension (OR, 0.50; 95% CI, 0.42-0.60; P corrected = 8.96 × 10−11), preoperative ambulation with assistance (OR, 1.75; 95% CI, 1.53-2.01; P corrected = 8.39 × 10−12), quitting smoking before the surgery (OR, 0.72; 95% CI, 0.61-0.85; P corrected = .05), and having a preoperative computed tomography angiography available (OR, 1.28; 95% CI, 1.16-1.43; P corrected = .02). Model characteristics for each of the three outcome are listed in Table I. The model for 30-day readmission had an area under the curve of 0.79 (95% CI, 0.74-0.82) (Fig). We developed a model to predict 30-day survival in claudicants undergoing infrainguenal bypass using clinical characteristics commonly available to physicians. This computational machine learning model was developed using real world VQI data and can possibly be used in a general clinic setting for preoperative risk stratification.TableModel characteristics used to predict 30-day survival, major adverse limb event (MALE) created from a composite of any limb amputation and any return to the operating room for revision, and major cardiac event (MCE) created from a composite of stroke, myocardial infarction, or death30-day survivalMALEMCEAUC (95% CI)0.79 (0.74-0.82)0.68 (0.57-0.72)0.71 (0.69-0.74)TPR (IQR)0.79 (0.66-0.91)0.55 (0.27-0.83)0.76 (0.50-0.90)FPR (IQR)0.46 (0.20-0.73)0.39 (0.10-0.69)0.47 (0.22-0.73)Cutoff0.10.030.1Sensitivity0.810.680.65Specificity0.640.510.65AUC, Area under the curve; CI, confidence interval; FPR, false positive rate; IQR, interquartile range; TPR, true positive rate. Open table in a new tab