Predicting Complications in Breast Reconstruction: Development and Prospective Validation of a Machine Learning Model.

Abstract

Necrosis of the nipple-areolar complex (NAC) is the Achilles heel of nipple-sparing mastectomy (NSM), and it can be difficult to assess which patients are at risk of this complication (Ann Surg Oncol 2014;21(1):100-106). To develop and validate a model that accurately predicts NAC necrosis in a prospective cohort. Data were collected from a retrospectively reviewed cohort of patients who underwent NSM and immediate breast reconstruction between January 2015 and July 2019 at our institution, a high -volume, tertiary academic center. Preoperative clinical characteristics, operative variables, and postoperative complications were collected and linked to NAC outcomes. These results were utilized to train a random-forest classification model to predict necrosis. Our model was then validated in a prospective cohort of patients undergoing NSM with immediate breast reconstruction between June 2020 and June 2021. Model predictions of NAC necrosis in the prospective cohort achieved an accuracy of 97% (95% confidence interval [CI], 0.89-0.99; P = 0.009). This was consistent with the accuracy of predictions in the retrospective cohort (0.97; 95% CI, 0.95-0.99). A high degree of specificity (0.98; 95% CI, 0.90-1.0) and negative predictive value (0.98; 95% CI, 0.90-1.0) were also achieved prospectively. Implant weight was the most predictive of increased risk, with weights greater than 400 g most strongly associated with NAC ischemia. Our machine learning model prospectively predicted cases of NAC necrosis with a high degree of accuracy. An important predictor was implant weight, a modifiable risk factor that could be adjusted to mitigate the risk of NAC necrosis and associated postoperative complications.