Predicting complications in breast reconstruction: External validation of a machine learning model

Abstract

Background

Nipple-sparing mastectomy (NSM) with immediate implant-based breast reconstruction provides aesthetic and psychosocial benefits, but nipple-areolar complex (NAC) necrosis remains a significant risk. This study externally validated a previously developed machine learning (ML) model that predicted NAC necrosis with 97% accuracy on institutional data.

Methods

This retrospective cohort study identified an initial 394 patients who underwent NSM with immediate breast reconstruction at Asan Medical Center (January 2021 - August 2022). Though there were 6 incomplete patient profiles, which resulted in 388 cases being used for statistical analysis, as demonstrated in Table 1. Demographic, oncologic, and surgical data were collected, with complications defined as post-operative events occurring within 90 days. A previously validated random forest ML model was applied to predict NAC necrosis. Model performance was assessed using accuracy, area under the receiver operating characteristic curve (AUC-ROC), sensitivity, specificity, and predictive values.

Results

Of 388 patients, 19 (4.9%) developed NAC necrosis. Significant risk factors included older age (mean age: 51.3 vs. 46.2 years, p = 0.015), higher BMI (mean BMI: 24.2 vs. 22.4, p = 0.024), active smoking (p = 0.008), and cumulative smoking exposure (mean pack-years: 3.3 vs. 0.2, p < 0.0001). Mastectomy specimen weight was significantly associated with NAC necrosis (mean: 394.2 g vs. 313.4 g, p = 0.021). The ML model achieved a predictive accuracy of 96%, with an AUC-ROC of 0.70 (95% CI: 0.55–0.85), indicating moderate discriminative ability.

Conclusions

The externally validated ML model accurately predicted NAC necrosis in a distinct patient population, demonstrating its potential for personalized risk assessment in NSM candidates. Future validation in diverse populations is needed.