186. Predicting Thromboembolic Events in Autologous Breast Reconstruction

PURPOSE: Breast cancer remains the most common malignancy among U.S. women, and autologous breast reconstruction (ABR) plays a pivotal role in the reconstructive process. Venous thromboembolism (VTE) is a serious complication following ABR, potentially leading to significant morbidity. This study aims to develop a machine learning model to predict VTE risk, with the goal of improving preoperative counseling and postoperative management, ultimately optimizing patient safety and outcomes. METHODS: A retrospective review of the American College of Surgeons National Surgical Quality Improvement Program (NSQIP) was used to identify patients who had undergone ABR from 2018-2022. Two machine learning methodologies were developed to predict VTE following ABR. Method I included all preoperative attributes and operative details, while Method II focused on variables statistically significant in univariate analysis. Five algorithms were trained: logistic regression, support vector machine, k-nearest neighbors, decision tree, and random forest. A neural network with four dense layers was also developed. Hyperparameters were optimized using grid search with 5-fold cross-validation, addressing class imbalance with SMOTE. RESULTS: In a cohort of 6,529 patients who underwent autologous breast reconstruction, 1.4% developed VTE. Pulmonary embolism occurred in 0.7%, and patients with VTE had significantly higher BMIs (p = 0.00004353) and longer operative times (p = 0.000165). These patients also had increased rates of reoperation (p < 0.0001) and unplanned readmissions (p < 0.0001). Postoperative complications like myocardial infarction (p = 0.007774) and septic shock (p = 0.000207) were also more prevalent in the VTE group. Method I identified the random forest, logistic regression, and k-nearest as top performing models, with random forest demonstrating the highest area under the receiver operating curve (0.99). Interestingly, the highest performing model with Method II was also the random forest with an area under the receiver operating curve of 0.87. CONCLUSION: The development of machine learning models to predict VTE following ABR offers a promising approach to improving patient safety and postoperative outcomes. Random forest, the top-performing algorithm in both methodologies, demonstrated high accuracy in identifying patients at risk of VTE, highlighting its potential to guide personalized preoperative counseling and interventions. This model could be further integrated into clinical practice to refine risk stratification, ultimately reducing the morbidity associated with VTE in this patient population. Future studies should focus on external validation and prospective application to maximize clinical utility.