临床淋巴结阴性早期肺腺癌隐匿性淋巴结转移的术前预测:一项多中心机器学习研究
本回顾性多中心研究纳入752例cN0早期肺腺癌患者,整合CT语义特征与影像组学特征构建可解释机器学习模型预测隐匿性淋巴结转移。XGBoost模型内部验证ROC-AUC为0.814,外部验证ROC-AUC为0.826,校准良好;风险分层显示从极低危至极高危组OLNM发生率单调递增。SHAP分析显示实变水平、Rad-score和分叶征为前三位预测因子。
要点
- XGBoost模型外部验证ROC-AUC为0.826(0.746-0.897),可用于术前预测cN0早期肺腺癌OLNM。
- 风险分层五档从极低危至极高危,OLNM发生率单调递增,内部验证为3.7%至40.0%。
- SHAP分析显示实变水平、影像组学Rad-score和分叶征为前三位预测因子。
核心结论
- 该验证模型整合医师判读的语义特征与数据驱动的影像组学,为个性化手术规划提供无创工具,可在资源有限环境中辅助制定淋巴结清扫策略。
研究设计与样本
- 回顾性多中心研究,共752例患者(训练集495例、内部验证集124例、外部验证集133例),采用ComBat协调扫描仪变异、PCA降维和LASSO回归筛选特征,SMOTE平衡训练数据后优化七种分类器。
诊断标准与关键数值
- XGBoost模型内部验证ROC-AUC为0.814(0.666-0.925),PR-AUC为0.502(0.328-0.803);外部验证ROC-AUC为0.826(0.746-0.897),PR-AUC为0.486(0.331-0.714);Brier评分分别为0.135和0.128。
临床应用与阅片要点
- 模型输出五档风险分层(极低危至极高危),对应OLNM发生率在内部验证中从3.7%升至40.0%,外部验证中极高危组转移风险为极低危组的4.1倍,可辅助术前评估淋巴结清扫范围。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
BACKGROUND: Occult lymph node metastasis (OLNM) in cN0 early-stage lung adenocarcinoma (LUAD) leads to pathological understaging and suboptimal surgical management. Current prediction tools exhibit limited robustness. This study aimed to develop and validate a machine learning model that integrates CT semantic and radiomic features for the preoperative prediction of OLNM.
METHODS: In this retrospective multicenter study, an interpretable machine learning model was developed. A cohort of 752 patients (training: n = 495; internal validation: n = 124; external validation: n = 133) underwent rigorous feature processing: ComBat harmonization for scanner variability, PCA dimensionality reduction, and LASSO regression for feature selection. Seven classifiers were optimized using SMOTE-balanced training data.
RESULTS: The XGBoost model demonstrated robust performance, achieving an ROCAUC of 0.814 (0.666-0.925) and a PR-AUC of 0.502 (0.328-0.803) in the internal validation cohort. It maintained strong generalizability in the external validation cohort, with an ROC-AUC of 0.826 (0.746-0.897) and a PR-AUC of 0.486 (0.331-0.714). The model was well-calibrated (Brier scores: 0.135 and 0.128, respectively). Risk stratification identified five clinically actionable tiers: "very-low-risk" to "very-high-risk" patients exhibited monotonically increasing rates of OLNM (internal validation: 3.7% to 40.0%; external validation: 4.1-fold increase in metastasis). SHAP analysis identified consolidation level, radiomics-derived Rad-score, and lobulation as the top three predictors.
CONCLUSION: This validated model integrates physician-interpreted semantics with data-driven radiomics, providing a non-invasive tool for personalized surgical planning. It enables tailored lymph node dissection strategies while enhancing accessibility in resource-limited settings.
DATA AVAILABILITY: The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.
引用
- Zhao F, Zhao Y, Ye Z, et al. Preoperative Prediction of Occult Lymph Node Metastasis in Clinically Node-Negative Early-Stage Lung Adenocarcinoma: A Multicenter Machine Learning Study. Acad Radiol. 2025. PMID: 41241594. doi:10.1016/j.acra.2025.10.058
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。