Using machine learning to predict outcomes following carotid endarterectomy

Objective Prediction of outcomes following carotid endarterectomy (CEA) remains challenging, with a lack of standardized tools to guide perioperative management. We used machine learning (ML) to develop automated algorithms that predict outcomes following CEA. Methods The Vascular Quality Initiative (VQI) database was used to identify patients who underwent CEA between 2003 and 2022. We identified 71 potential predictor variables (features) from the index hospitalization (43 preoperative [demographic/clinical], 21 intraoperative [procedural], and 7 postoperative [in-hospital complications]). The primary outcome was stroke or death at 1 year following CEA. Our data were split into training (70%) and test (30%) sets. Using 10-fold cross-validation, we trained six ML models using preoperative features (Extreme Gradient Boosting [XGBoost], random forest, Naïve Bayes classifier, support vector machine, artificial neural network, and logistic regression). The primary model evaluation metric was area under the receiver operating characteristic curve (AUROC). After selecting the best performing algorithm, additional models were built using intra- and postoperative data. Model robustness was evaluated using calibration plots and Brier scores. Performance was assessed on subgroups based on age, sex, race, ethnicity, insurance status, symptom status, and urgency of surgery. Results Overall, 166,369 patients underwent CEA during the study period. In total, 7749 patients (4.7%) had the primary outcome of stroke or death at 1 year. Patients with an outcome were older with more comorbidities, had poorer functional status, and demonstrated higher risk anatomic features. They were also more likely to undergo intraoperative surgical re-exploration and have in-hospital complications. Our best performing prediction model at the preoperative stage was XGBoost, achieving an AUROC of 0.90 (95% confidence interval [CI], 0.89-0.91). In comparison, logistic regression had an AUROC of 0.65 (95% CI, 0.63-0.67), and existing tools in the literature demonstrate AUROCs ranging from 0.58 to 0.74. Our XGBoost models maintained excellent performance at the intra- and postoperative stages, with AUROCs of 0.90 (95% CI, 0.89-0.91) and 0.94 (95% CI, 0.93-0.95), respectively. Calibration plots showed good agreement between predicted and observed event probabilities with Brier scores of 0.15 (preoperative), 0.14 (intraoperative), and 0.11 (postoperative). Of the top 10 predictors, eight were preoperative features, including comorbidities, functional status, and previous procedures. Model performance remained robust on all subgroup analyses. Conclusions We developed ML models that accurately predict outcomes following CEA. Our algorithms perform better than logistic regression and existing tools, and therefore, have potential for important utility in guiding perioperative risk mitigation strategies to prevent adverse outcomes.