Objectives: Papillary thyroid carcinoma (PTC) frequently presents with cervical lymph node metastasis (CLNM), yet preoperative tools often encode BRAF V600E as a binary variable, potentially overlooking information contained in mutation abundance. We sought to quantify the dose–response relationship between BRAF V600E abundance and CLNM and to develop an interpretable model for preoperative risk stratification.Methods: We performed a single-center retrospective study of consecutive PTC patients who underwent preoperative BRAF V600E testing and surgery from 2019 to 2023. Patients were randomly split 70/30 into training and test sets. Candidate predictors included clinical and ultrasound features and BRAF V600E abundance. We used multivariable logistic regression and restricted cubic splines (RCS) to assess nonlinearity and compared six machine-learning algorithms (LR, KNN, SVM, XGB, LightGBM, and NN). Model performance was evaluated by F1, AUC, calibration, and decision-curve analyses; SHAP aided interpretation. Ethics approval: SYSKY-2024-169-01.Results: The cohort included 667 patients; CLNM occurred in 391 (58.6%). CLNM cases had higher BRAF abundance (median 23% vs. 17%) and characteristic clinical/sonographic differences. RCS revealed a nonlinear association between abundance and CLNM, with a steep risk rise of up to ~20.7% followed by a plateau. Among six algorithms, XGBoost showed the best validation performance (AUC 0.752; F1 0.73). SHAP indicated that maximum tumor diameter, BRAF abundance, age, and microcalcifications contributed most to predictions.Conclusions: Modeling BRAF V600E as a quantitative abundance—rather than a binary status—improves preoperative CLNM risk assessment in PTC. An interpretable XGBoost model integrating abundance with routine features demonstrates acceptable discrimination and potential clinical utility for individualized surgical planning and counseling.
目的：甲状腺乳头状癌（PTC）常伴有颈部淋巴结转移（CLNM），但术前检查工具常将BRAF V600E编码为二元变量，可能忽略了突变丰度所包含的信息。本研究旨在量化BRAF V600E丰度与CLNM之间的剂量-反应关系，并建立一个可解释的模型用于术前风险分层。 方法：我们对2019年至2023年间连续接受术前BRAF V600E检测和手术的PTC患者进行了单中心回顾性研究。患者按7:3的比例随机分为训练集和测试集。候选预测因子包括临床特征、超声特征及BRAF V600E丰度。我们采用多变量逻辑回归和限制性立方样条（RCS）评估非线性关系，并比较了六种机器学习算法（逻辑回归、K近邻、支持向量机、XGBoost、LightGBM和神经网络）。通过F1分数、AUC、校准曲线和决策曲线分析评估模型性能；SHAP方法辅助模型解释。伦理批准号：SYSKY-2024-169-01。 结果：队列共纳入667例患者，其中391例（58.6%）发生CLNM。CLNM患者的BRAF丰度更高（中位数23% vs. 17%），并具有特征性的临床及超声差异。RCS分析显示丰度与CLNM之间存在非线性关联，风险在丰度达到约20.7%前急剧上升，随后进入平台期。在六种算法中，XGBoost在验证集上表现出最佳性能（AUC 0.752；F1分数 0.73）。SHAP分析表明，最大肿瘤直径、BRAF丰度、年龄和微钙化对预测贡献最大。 结论：将BRAF V600E作为定量丰度（而非二元状态）进行建模，可改善PTC术前CLNM风险评估。一个整合了丰度与常规特征的可解释XGBoost模型显示出可接受的区分度，在个体化手术规划和患者沟通方面具有潜在的临床应用价值。
肿瘤(癌症)患者之家