ARRS 2022 Abstracts


2074. An Interpretable Machine Learning Model for Prediction of 30-day Amputation Events Following Lower Extremity Endovascular Procedures
Authors * Denotes Presenting Author
  1. Meredith Cox *; Massachusetts General Hospital
  2. Nicholas Reid; Massachusetts General Hospital
  3. J.C. Panagides; Massachusetts General Hospital
  4. Sanjeeva Kalva; Massachusetts General Hospital
  5. Jayashree Kalpathy-Cramer; Massachusetts General Hospital
  6. Dania Daye; Massachusetts General Hospital
Severe peripheral artery disease (PAD) may result in lower extremity amputation or require multiple surgical or endovascular interventions to achieve limb salvage. Current prediction models for major amputation risk have had limited performance at the individual level. The goal of this study is to build a machine learning model that may enhance short-term major amputation outcome prediction in patients with PAD undergoing endovascular interventions to aid in improved clinical decision-making.

Materials and Methods:
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 laboratory information on the 14,444 patients who underwent lower extremity endovascular procedures for PAD between 2011 and 2018. Our primary outcome of interest was 30-day ipsilateral major amputation of the affected limb. Missing values were imputed using Optimal Imputation to avoid bias caused by excluding cases with incomplete data. Features were selected using Minimum Redundancy Maximum Relevance (mRMR) with an incremental feature selection (IFS) approach, yielding an optimal feature set of 12 variables. Using data from 2011-2017 for training and data from 2018 for validation, we developed and tested a machine learning model to predict the occurrence of 30-day amputation following infra-inguinal endovascular procedures. We subsequently interpreted the model using Gini Importance metrics and Shapley Additive exPlanations (SHAP). We also calculated performance metrics across different races, sexes, and age groups to ensure model fairness.

The machine learning model built using a random forest algorithm achieved an AU-ROC of 0.81. The most important features of the model were elective surgery designation, white blood cell count, claudication, albumin, and hematocrit. The model performed equally well on white and nonwhite patients (Delong p-value = 0.903), males and females (Delong p-value = 0.649), and patients younger than 65 and patients age 65 and older (Delong p-value = 0.792).

We present a machine learning model that predicts 30-day major amputation events in PAD patients undergoing lower extremity endovascular procedures which is robust across different races, sexes, and age groups. This model can aid in optimizing clinical decision-making in patients with PAD.