Validating machine learning approaches for prediction of donor related complication in microsurgical breast reconstruction: a retrospective cohort study

Abstract

Abstract Autologous reconstruction using abdominal flaps remains the most popular method for breast reconstruction worldwide. We aimed to evaluate a prediction model using machine-learning methods and to determine which factors increase abdominal flap donor site complications with logistic regression. We evaluated the predictive ability of different machine learning packages, reviewing a cohort of breast reconstruction patients who underwent abdominal flaps. We analyzed 13 treatment variables for effects on the abdominal donor site complication rates. To overcome data imbalances, random over sampling example (ROSE) method was used. Data were divided into training and testing sets. Prediction accuracy, sensitivity, specificity, and predictive power (AUC) were measured by applying neuralnet, nnet, and RSNNS machine learning packages. A total of 568 patients were analyzed. The supervised learning package that performed the most effective prediction was neuralnet. Factors that significantly affected donor-related complication was size of the fascial defect, history of diabetes, muscle sparing type, and presence or absence of adjuvant chemotherapy. The risk cutoff value for fascial defect was 37.5 cm2. High-risk group complication rates analyzed by statistical method were significant compared to the low-risk group (26% vs 1.7%). These results may help surgeons to achieve better surgical outcomes and reduce postoperative burden.