(IP255) Using Machine Learning Algorithms to Predict the Safest and Most Effective Carotid Revascularization Procedure Based on Patient Specific Factors

Objectives: Carotid endarterectomy (CEA), transcarotid artery revascularization (TCAR), and transfemoral carotid artery stenting (TFCAS) are all currently approved for stroke prevention in patients with carotid artery occlusive disease. Existing outcome prediction tools are limited to simple regression models for a single procedure or short-term outcomes. Therefore, we developed a comprehensive machine learning algorithm encompassing all three procedures to predict the risk of one-year stroke/death, to recommend the safest and most effective procedure tailored for each patient’s own specific anatomy and baseline characteristics.

Methods: The Vascular Quality Initiative database was queried for patients undergoing CEA, TCAR, or TFCAS from 2016 onwards. The primary outcome was one-year stroke/death. Utilizing four common machine learning (ML) algorithms – generalized linear model, gradient boosted machine (GBM), neural network, and random forest- we trained 20 models per algorithm, alongside a stacked ensemble model that integrated the best-performing base models across all algorithms. We developed a surrogate approach that translated complex GBM predictions into procedure-specific risk scoring systems. Over 37 variables were analyzed [demographics, anatomy (lesion, neck, arch), prior procedures, comorbidities, medications]. The fitted model coefficients were converted into weighted point values to determine the predictors of 1-year stroke/death to quantify the risk of this outcome for each of the three procedures.

Results: Our study included 239,515 patients [CEA 142,137 (59%), TCAR 66,655 (28%), TFCAS 30,723 (13%)]. The dataset was split into training (70%) and test (30%) subsets. The GBM model achieved an area under the receiver operating characteristic curve (AUC) of 0.74. The surrogate scoring system demonstrated strong correlation (r=0.9) with the GBM’s predicted risks based on randomly selected sample of patients, validating its use as a simplified clinical tool. Factors mostly predicting one-year stroke/death included age, sex, congestive heart failure (CHF), diabetes, and symptomatic status (Table I). Table II show a real-world example of predicting the safest procedure for three different example patients based on their own specific data and surrogate scores.

Conclusions: To our knowledge, this is the first ML prediction model to include CEA, TCAR and TFCAS utilizing the largest number of patients ever recorded with prediction of one-year outcomes. Our surrogate scoring system effectively predicts the likelihood of one-year stroke/death across the three procedures. This provides a tool for risk stratification while maintaining strong correlation with the sophisticated ML predictions. The integrated ML predictive framework aids in providing patient specific recommendations based on each procedure safety and effectiveness in preventing one-year stroke.