Predicting Nipple–Areolar Complex Necrosis in Nipple-sparing Mastectomy with a Machine Learning Model

Abstract

BACKGROUND & OBJECTIVES: Necrosis of the nipple-areolar complex (NAC) is the Achilles’ heel of Nipple Sparing Mastectomy (NSM).1 Our goal was to create a user-friendly, validated technology platform that surgeons could utilize preoperatively to predict which patients are at risk of NAC necrosis and aid in shared decision-making. This prediction would be based on patient characteristics and intraoperative variables. METHODS: We conducted a retrospective review of all NSM with immediate reconstruction performed at our institution between January 2015 and July 2019. Preoperative clinical characteristics, operative variables, and postoperative complications were collected and linked to NAC outcomes. These results were utilized to train a Random Forest machine learning model to predict necrosis when given preoperative patient characteristics, including age, BMI, smoking status, medical history, previous breast incisions, breast size, and planned implant or tissue expander (TE) size. Our model was subsequently validated by predicting NAC outcomes in a prospective cohort. The prospective cohort included all therapeutic and prophylactic NSM performed at our institution from May 2020 through October 2020. RESULTS: 305 breasts in 181 patients were included in the retrospective dataset which served as the foundation for our machine learning model. Full-thickness skin necrosis including part or all of the NAC occurred in 46 cases (15.1%). The strongest predictors of NAC necrosis were implant weight (P < 0.001) and weight of the mastectomy specimen (P < 0.001). When controlling for implant volume in our model, fill weight maintained a strong association with NAC necrosis. Rates of NAC necrosis among TE reconstructions using air only were lower at equivalent volumes compared with TE using saline fill and direct to implant reconstructions. Other influential factors included diabetes (P = 0.005), smoking (P = 0.04), and hypertension (= 0.03). With a prospective cohort of 27 patients, our predictive machine learning model achieved 96% accuracy (P = 0.02) with high positive predictive value and specificity for NAC necrosis. The model correctly predicted 4 of 5 cases of NAC necrosis and all 22 cases without necrosis. CONCLUSIONS: Implant weight is an independent risk factor for NAC necrosis following NSM, indicating that lower implant volumes, or using air-only initial TE fill, may mitigate the risk of NAC necrosis. The findings of our predictive Random Forest machine learning model also provide a basis for utilizing artificial intelligence to predict cases with a high probability of NAC necrosis. We created an easy-to-use interface for our model, which allows a user to input patient characteristics and receive a prediction which includes a binary outcome: “NAC Necrosis” or “No NAC Necrosis” and the predicted probability of necrosis. The instrument also provides a description of each variable’s effect on the prediction. Such models may be developed using institutional data and utilized to inform patient decision-making prior to mastectomy. REFERENCE: 1. Wapnir I, Dua M, Kieryn A, et al. Intraoperative imaging of nipple perfusion patterns and ischemic complications in nipple-sparing mastectomies. Ann Surg Oncol. 2014;21(1):100–106.