AI Risk Prediction Tools for Alloplastic Breast Reconstruction

INTRODUCTION: Accurate risk prediction for patients undergoing breast reconstruction with tissue expanders (TEs) can improve patient counseling and shared decision-making. This study aimed to develop and evaluate traditional statistical and machine learning (ML) approaches to predicting complications in alloplastic breast reconstruction. METHODS: Patient characteristics, surgical techniques, and complications were collected for all women undergoing immediate TE placement from 2017-2023 at Memorial Sloan Kettering Cancer Center. Multivariable logistic regression and ML models were developed to predict TE loss, infection, and seroma. ML model performance was optimized using ten-fold cross validation with hyperparameter tuning. Evaluation metrics included area under the receiver operating curve (AUC), sensitivity, specificity, and Brier score. RESULTS: This study included 4,046 women undergoing 6,513 immediate TE placements. TE loss occurred in 7.6% of patients (4.8% of TEs), infection in 10% of patients (7.2% of TEs), and seroma in 11.5% of patients (6.2% of TEs). Traditional multivariable regression demonstrated AUCs of 0.63-0.69 and ML models demonstrated AUCs of 0.71-0.73 in predicting TE complications. SHAP analysis highlighted BMI, prepectoral placement, and chemotherapy as key predictors of TE complications. Top-performing models were built into nomograms and a web-based prediction application to provide real-time risk estimates based on patient-specific information. CONCLUSION: Accurate risk prediction tools using nomograms and ML models were developed to predict complications in alloplastic breast reconstruction. These findings support incorporating both traditional statistics and machine learning analyses into preoperative assessments of patients undergoing alloplastic breast reconstruction to enhance data-driven, personalized care.