IYSC02 - Machine Learning Model of Clinical Risk Factors and Peripheral Calcium Scoring on Computed Tomography for the Prediction of Amputation in Patients with Peripheral Arterial Disease

Objectives: To date, there is no quantifiable evidence of the preferred measure of peripheral arterial disease (PAD) to best predict outcomes. Lower extremity calcium scoring is a novel tool that helps risk stratify patients with PAD. This study uses a machine learning-based approach to develop risk scores based on clinical variables and peripheral calcium scoring in an attempt to predict risk of major amputation in patients with PAD.

Methods: A multi-institutional database was queried for patients with diagnosis of PAD who had undergone multidetector computed tomography of the aorta and bilateral lower extremity runoffs from 2016 and 2020. Patients that underwent prior balloon angioplasty without stenting or atherectomy were included; those with prior surgical (open bypass) or endovascular intervention beyond angioplasty (atherectomy, stenting), were excluded. The dataset was then split into a training and test dataset with a 3:1 ratio. ML models were constructed using optimal trees with Bayesian optimization to predict the risk of amputation using maximum relevance – minimum redundancy (MRMR) feature selection to identify the best performing classifiers. Performance comparison of the ML algorithm was evaluated with the receiver operating characteristic curve. The model was then validated using the test dataset.

Results: A total of 142 patients were identified and included in the dataset, 13 of whom underwent amputation. Based on the MRMR feature selection, the top 15 variables included in the ML model were the following: history of stroke, diabetes, HLD, CKD, smoker, presence of lower extremity wound infection, history of endovascular reinterventions, history of surgical bypass, and calcium score in the anterior tibial, peroneal, and external iliac arteries. Performance in predicting amputation on the validation set had an accuracy of 94.3% with AUC of 0.79, sensitivity of 33%, specificity of 100%. The negative predictive value was 94%, the positive predictive value 100%. The predictors with the highest Shapley values were in order the following: number of endovascular reinterventions, history of diabetes, and external iliac artery calcium score.

Conclusions: A machine learning model utilizing lower extremity calcium scores can help identify PAD patients with a high risk for amputation.