基于人工智能的合并慢性淋巴细胞性甲状腺炎的甲状腺乳头状癌颈中央区淋巴结转移的早期预测系统构建

Construction of early prediction system of central cervical lymph node metastasis of papillary thyroid carcinoma complicated with chronic lymphocytic thyroiditis based on artificial intelligence

目的分析合并慢性淋巴细胞性甲状腺炎(chronic lymphocytic thyroiditis,CLT)的甲状腺乳头状癌(papillary thyroid carcinoma,PTC)患者发生颈中央区淋巴结转移(central lymph node metastasis,CLNM)的危险因素,并基于机器学习算法建立预测系统。方法回顾性选取2018年5月~2020年12月于南京大学医学院附属南京鼓楼医院接受手术治疗并经病理确诊为PTC且合并CLT的患者429例。采用Logistic回归模型筛选CLNM发生的危险因素,分别建立CatBoost、XGBoost及LightGBM预测模型。结果年龄、肿瘤直径、病灶数目及腺外侵犯是合并CLT的PTC患者发生CLNM的预测因素(P<0.05)。训练集中,3种模型的AUC分别为0.884(95%CI 0.836-0.932)、0.811(95%CI 0.748-0.874)、0.816(95%CI 0.749-0.883),准确度分别为92.33%、83.67%、85.00%;测试集中,3种模型的AUC分别为0.852(95%CI 0.795-0.909)、0.803(95%CI 0.742-0.865)、0.795(95%CI 0.726-0.864),准确度分别为91.47%、80.62%、83.72%。结论基于机器学习建立的3种预测模型中,CatBoost模型的性能更优,在CLT合并PTC中预测CNM的发生具有更好的应用价值。

OBJECTIVE To analyze the risk factors of central lymph node metastasis(CLNM)in papillary thyroid carcinoma(PTC)patients with chronic lymphocytic thyroiditis(CLT),and to establish a prediction system based on machine learning algorithm.METHODS We retrospectively selected 429 patients who underwent surgery in our hospital from May 2018 to December 2020 and were pathologically diagnosed with PTC and CLT.Logistic regression model was used to screen the risk factors of CLNM,and CatBoost,XGBoost and LightGBM prediction models were established respectively.RESULTS Age,tumor diameter,number of lesions,and extrathyroidal extension were predictors of CLNM in PTC patients with CLT(P<0.05).In the training set,the AUC of the three models were 0.884(95%CI 0.836-0.932),0.811(95%CI 0.748-0.874),0.816(95%CI 0.749-0.883),and the accuracy was 92.33%,83.67%,85.00%respectively.In the test set,the AUC of the three models were 0.852(95%CI 0.795-0.909),0.803(95%CI 0.742-0.865),0.795(95%CI 0.726-0.864),and the accuracy was 91.47%,80.62%,83.72%respectively.CONCLUSION Among the three prediction models based on machine learning algorithm,CatBoost model has better performance and better application value in predicting CNM occurrence in CLT combined with PTC.