基于机器学习的保留乳头乳房切除术中乳头–乳晕复合体受累预测模型的开发与验证

Development and Validation of a Machine Learning Based Predictive Model for Nipple-Areolar Complex Involvement in Nipple-Sparing Mastectomy

摘要: 目的：开发并验证一个整合临床特征和机器学习算法的预测模型，用于评估拟行保留乳头乳房切除术患者中乳头–乳晕复合体的受累情况。方法：这项回顾性研究分析了青岛大学附属医院接受NSM并同期乳房重建的238例乳腺癌患者。根据术后病理，患者被分为NAC受累组(n = 36)和NAC未受累组(n = 202)。评估了临床、影像学和病理学特征。进行了单因素和多因素分析，并开发比较了四种机器学习模型——逻辑回归、随机森林、支持向量机和XGBoost。结果：乳头内陷、血性乳头溢液、肿瘤–乳头距离 ≤ 2 cm、淋巴血管浸润和临床淋巴结状态被确定为NAC受累的独立预测因子(P < 0.05)。XGBoost模型表现出最高的预测性能，曲线下面积为0.977，其次是随机森林(AUC = 0.970)、逻辑回归(AUC = 0.968)和SVM (AUC = 0.958)。基于显著预测因子的整合模型在基于逻辑回归的列线图验证中显示出0.775的AUC。多因素分析证实乳头内陷和血性乳头溢液是强有力的独立危险因素(OR = 12.83, OR = 18.64, P < 0.05)。结论：结合关键临床预测因子与机器学习建模，为术前预测NSM候选者NAC受累提供了一个有价值的工具。该方法有助于早期识别高危患者，可能优化手术规划和肿瘤学安全性。

Abstract: Objective: To develop and validate a predictive model incorporating clinical features and machine learning algorithms for assessing nipple-areolar complex (NAC) involvement in patients undergoing nipple-sparing mastectomy (NSM). Methodology: This retrospective study analyzed 238 patients with breast cancer who underwent NSM with immediate breast reconstruction at the Affiliated Hospital of Qingdao University. Based on postoperative pathology, patients were categorized into an NAC-involved group (n = 36) and an NAC-not-involved group (n = 202). Clinical, imaging, and pathological features were evaluated. Univariate and multivariate analyses were performed, and four machine learning models—Logistic Regression, Random Forest, Support Vector Machine (SVM), and XGBoost—were developed and compared. Results: Nipple retraction, bloody nipple discharge, tumor-to-nipple distance (TND) ≤ two cm, lymphovascular invasion, and clinical lymph node status were identified as independent predictors of NAC involvement (P < 0.05). The XGBoost model demonstrated the highest predictive performance, with an area under the curve (AUC) of 0.977, followed by Random Forest (AUC = 0.970), Logistic Regression (AUC = 0.968), and SVM (AUC = 0.958). The integrated model based on significant predictors showed an AUC of 0.775 in logistic regression-based nomogram validation. Multivariate analysis confirmed nipple retraction and bloody nipple discharge as strong independent risk factors (OR = 12.83 and OR = 18.64, P < 0.05). Conclusion: The combination of key clinical predictors and machine learning modeling provides a valuable tool for preoperative prediction of NAC involvement in NSM candidates. This approach aids in the early identification of high-risk patients, potentially optimizing surgical planning and oncologic safety.