Abstract: The clinical value of metastasis of lymph node posterior to right recurrent laryngeal nerve in papillary thyroid carcinoma based on machine learning algorithm        ZHOU Tian-han*，WU Fan，LU Kai-ning，et al. *The Fourth School of Clinical Medicine, Zhejiang Chinese Medical University, Hangzhou310006, China
Corresponding author：LUO Ding-cun，E-mail:ldc65@163.com
Abstract    Objective    To construct and verify a prediction model based on machine learning algorithm for the metastasis of lymph node posterior to right recurrent laryngeal nerve（LN-prRLN）in papillary thyroid carcinoma（PTC）.Methods    A retrospective survey was conducted on 907 PTC patients who underwent surgery in the Department of Surgical Oncology of Hangzhou First People's Hospital from March 2014 to July 2019. The clinicopathological data of gender, age, tumor size, membranous invasion and multifocal tumor were included in this study. According to the time series，they were divided into training set（n=595）and validation set（n=312）.Logistic regression and machine learning algorithms such as classification tree，random forest，gradient boosting machine and support vector machine were used to select characteristic variables，and the prediction model of metastasis of LN-prRLN was constructed. Results    The area under the receiver operating characteristic curve（ROC）of the classification tree model was 0.654，the sensitivity, specificity, accuracy were  33.00%, 97.78%, 86.89% respectively; The area under ROC curve of the random forest model was 0.753，the sensitivity, specificity, accuracy were 57.00% ,100.00% , 92.77% respectively. The area under ROC curve of support vector machine model was 0.604，the sensitivity, specificity, accuracy were 27.00% ，83.19% ，86.39% respectively. The area under the ROC curve of gradient boosting machine was 0.873，the sensitivity, specificity, accuracy were 72.00%, 89.49%, 87.90% respectively. Conclusion    The prediction model of metastasis LN-prRLN has a good prediction effect. The gradient boosting machine has the highest diagnostic efficiency.

Key words: lymph node posterior to right recurrent laryngeal nerve, papillary thyroid carcinoma, machine learning algorithm, prediction model

摘要： 目的    构建基于机器学习算法的甲状腺乳头状癌右喉返神经后方淋巴结（LN-prRLN）转移预测模型，并验证其预测效果。方法    回顾性分析2014年3月至2019年7月在浙江大学医学院附属杭州市第一人民医院肿瘤外科接受手术的907例甲状腺乳头状癌病人。分别纳入性别、年龄、肿瘤大小、被膜侵犯、多灶性等临床病理资料。根据时间序列，分为训练组（n=595）和验证组（n=312）。运用Logistic回归及分类树、随机森林、梯度提升法、支持向量机等机器学习算法进行特征变量选择，并构建LN-prRLN转移的预测模型。结果    分类树模型的受试者操作特征曲线（ROC）曲线下面积为0.654，敏感度为33.00%，特异度为97.78%，准确率为86.89%；随机森林模型的ROC曲线下面积为0.753，敏感度为57.00%，特异度为100.00%，准确率为92.77%；支持向量机模型的ROC曲线下面积为0.604，敏感度为27.00%，特异度为83.19%，准确率为86.39%；梯度提升法的ROC曲线下面积为0.873，敏感度为72.00%，特异度为89.49%，准确率为87.90%。结论    LN-prRLN转移预测模型对甲状腺乳头状癌右喉返神经后方淋巴结转移具有良好的预测效果，其中梯度提升法具有较高的诊断效能。

关键词: 右喉返神经后方淋巴结, 甲状腺乳头状癌, 机器学习算法, 预测模型