Advertisement

AI4 Comparing Machine-Learning Methods for the Prediction of Major Adverse Limb Events and Mortality after a Percutaneous Intervention

      Objectives

      The objective was to formulate, test, and compare the performance of regression-based and machine learning models in the prediction major adverse limb events (MALE) and mortality among patients receiving treatment for lower extremity peripheral artery disease (PAD).

      Methods

      Patients undergoing atherectomy, stent, and combination stent atherectomy for lower extremity PAD were identified in the Vascular Quality Initiative registry. Thirty-nine variables summarizing demographic, medical history, pre-operative, indication-specific, and procedure-specific characteristics were utilized to predict MALE and mortality events. For both events, we compared the performance of four different prediction models: a generalized linear model (GLM), a Least Absolute Shrinkage and Selection Operator (LASSO) regularized GLM, a gradient boosted decision tree, and random forest model. The area under the curve (AUC) evaluated the effectiveness of each prediction model. For validation purposes, 5-fold cross-validation was repeated three times. Pairwise comparisons of the receiver operating characteristic curves (ROC), sensitivity, and specificity measures with Bonferroni adjustment for multiple testing applied were performed to compare the models' performance.

      Results

      Among 15964 identified patients, a MALE occurred in 26.02% of patients, and death occurred in 18.82% of patients. The most effective predictive model for MALE, as determined by the AUC, was the gradient boosted decision tree (AUC= 0.7539) followed by the LASSO regulated GLM (AUC= 0.749). The most effective predictive model for mortality was the LASSO regularized GLM (AUC=0.7930) followed by the GLM model (AUC=0.7922). The GLM, LASSO regularized GLM model, and gradient boosted decision tree produced similar ROC.

      Conclusions

      All models showed acceptable discrimination, with an AUC greater than 0.7, when predicting MALE and mortality among patients receiving treatment for lower extremity peripheral artery disease. The machine learning techniques outperformed traditional regression-based techniques and can be leveraged to generate robust predictive models within the clinical space of lower extremity PAD.