机器学习:一项预测cN0甲状腺乳头状癌侧颈淋巴结转移的多中心研究
本研究基于医院A的1616例经细针穿刺活检诊断的甲状腺乳头状癌患者(按7:3分为训练集与测试集)及医院B的243例验证集,比较传统列线图与8种机器学习模型预测同侧侧颈及II、III、IV区淋巴结转移的性能;随机森林表现最佳,在测试集/验证集中预测同侧侧颈淋巴结转移的准确度、AUC、特异度、敏感度分别为0.773/0.728、0.858/0.799、0.984/0.935、0.757/0.807,并优于传统列线图。
要点
- 随机森林在测试集/验证集预测同侧侧颈淋巴结转移的AUC为0.858/0.799。
- 基于前10位贡献特征(含同侧中央区淋巴结转移率、甲状腺外侵犯、同侧中央区淋巴结转移数目)的简化模型性能仍强。
- 机器学习模型在多个指标及PMRA分析中清楚优于传统列线图。
核心结论
- 机器学习尤其是随机森林能可靠预测cN0甲状腺乳头状癌患者的侧颈淋巴结转移,优于传统列线图。
研究设计与样本
- 多中心研究,医院A的1616例经细针穿刺活检诊断的甲状腺乳头状癌患者按7:3分为训练集与测试集,医院B的243例患者作为验证集。
- 分析同侧侧颈及II、III、IV区淋巴结转移四个因变量,开发8种机器学习模型(逻辑回归、决策树、随机森林、梯度提升、支持向量机、K近邻、高斯朴素贝叶斯、神经网络),采用10折交叉验证和网格搜索超参数调优进行验证。
- 使用包括准确度、曲线下面积(AUC)、特异度和敏感度在内的11项指标评估模型,并以基于概率的排名模型方法(PMRA)将最佳模型与列线图比较。
主要发现
- 随机森林在测试集/验证集中预测同侧侧颈淋巴结转移的准确度、AUC、特异度、敏感度分别为0.773/0.728、0.858/0.799、0.984/0.935、0.757/0.807。
- 基于前10位贡献特征(包括同侧中央区淋巴结转移率、甲状腺外侵犯、同侧中央区淋巴结转移数目)的简化模型保持了强大性能,并在多项指标及PMRA分析中清楚超越传统列线图。
- 其他因变量获得类似结果,随机森林模型依赖不同但有重叠的特征集。
- 通过针对4个因变量各自的网络计算器促进临床工具实施。
关键数值与推荐
- 随机森林预测同侧侧颈淋巴结转移的AUC在测试集为0.858、验证集为0.799。
- 简化模型基于前10位贡献特征,包括同侧中央区淋巴结转移率、甲状腺外侵犯、同侧中央区淋巴结转移数目。
- 模型评估采用11项指标,包括准确度、AUC、特异度和敏感度。
临床怎么用
- 原文未据此给出推荐。
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)。
原文
BACKGROUND: The necessity of prophylactic lateral neck dissection for cN0 papillary thyroid carcinoma (PTC) remains debated. This study aimed to compare traditional nomograms with machine learning (ML) models for predicting ipsilateral lateral and level II, III, and IV lymph node metastasis (LNM). METHODS: Data from 1616 PTC patients diagnosed via fine-needle aspiration biopsy from hospital A were split into training and testing sets (7:3). Two hundred forty-three patients from hospital B served as a validation set. Four dependent variables-ipsilateral lateral and level II, III, and IV LNM-were analyzed. Eight ML models [logistic regression, decision tree, random forest (RF), gradient boosting, support vector machine, K-nearest neighbor, Gaussian naive Bayes, neural networks] were developed and validated using 10-fold cross-validation and grid search hyperparameter tuning. Models were assessed using 11 metrics including accuracy, area under the curve (AUC), specificity, and sensitivity. The best was compared with nomograms using the probability-based ranking model approach (PMRA). RESULTS: RF outperformed other approaches achieving accuracy, AUC, specificity, and sensitivity of 0.773/0.728, 0.858/0.799, 0.984/0.935, 0.757/0.807 in the testing/validation sets, respectively, for ipsilateral LLNM. A streamlined model based on the top 10 contributing features that includes ipsilateral central lymph node metastasis rate, extrathyroidal extension, and ipsilateral central lymph node metastasis number retained strong performance and clearly surpassed a traditional nomogram approach based on multiple metrics and PMRA analysis. Similar results were obtained for the other dependent variables, with the RF models relying on distinct but overlapping sets of features. Clinical tool implementation is facilitated via a web-based calculator for each of the 4 dependent variables. CONCLUSION: ML, especially RF, reliably predicts lateral LNM in cN0 PTC patients, outperforming traditional nomograms.
引用
- Zhou J, Li D, Ren J, et al. Machine Learning: A Multicenter Study on Predicting Lateral Lymph Node Metastasis in cN0 Papillary Thyroid Carcinoma. J Clin Endocrinol Metab. 2025. PMID: 39921476. doi:10.1210/clinem/dgaf070
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。