摘要
目的对比评估轻梯度提升(LightGBM)、自适应梯度提升(AdaBoost)、K近邻分类(KNN)及支持向量机(SVM)四种机器学习模型在女性甲状腺结节良恶性鉴别中的诊断性能,筛选最优预测模型。方法收集广西医科大学第八附属医院2023年1月1日至2024年1月31日间经病理确诊的447例女性甲状腺良恶性结节患者病例资料。通过单因素和多因素logistic回归分析筛选关键变量,将样本分为建模集(2023年1月—10月,329例)和独立外部验证集(2023年11月—2024年1月,118例);建模集进一步按7∶3随机划分为训练集与测试集,采用5折交叉验证评估模型性能。以敏感度、受试者工作特征曲线下面积(AUC)、校准性及临床决策曲线等指标评价模型鉴别效能。结果多因素logistic回归分析显示,超声特征(结节多灶性、纵横比、C-TIRADS分级)及实验室指标甲状腺过氧化物酶抗体(TPOAb)为关键预测变量(均P<0.05)。四种模型中,LightGBM表现最优:训练集AUC为0.978,验证集AUC为0.918,独立外部验证集AUC为0.889,且在校准性、临床决策曲线及敏感度方面均展现显著临床实用性。结论LightGBM模型可为临床早期鉴别女性甲状腺结节良恶性提供高效、稳定的辅助工具。
Objective To comparatively evaluate the diagnostic performance of four machine learning models(light gradient boosting machine[LightGBM],adaptive boosting[AdaBoost]and K-nearest neighbors[KNN],and support vector machine[SVM])in differentiating benign and malignant thyroid nodules in female patients,so as to identify the optimal predictive model.Methods Clinical data of 447 female patients with pathologically confirmed benign or malignant thyroid nodules were collected from the Eighth Affiliated Hospital of Guangxi Medical University—Guigang City People's Hospital between 1 January 2023 and 31 January 2024.Key predictive variables were screened using univariate and multivariate logistic regression analysis.The samples were divided into modeling set(n=329,collected from January 2023 to October 2023)and independent external validation set(n=118,collected from November 2023 to January 2024).The modeling set was further randomly split into a training set and a test set at a 7∶3 ratio.Model performance was validated using 5-fold cross-validation.The discriminative performance of the models was evaluated by indicators including sensitivity,area under the receiver operating characteristic curve(AUC),calibration,and clinical decision curve.Results Multivariate logistic regression analysis revealed that ultrasonographic features(multifocality,anteroposterior-to-transverse diameter ratio,and Chinese thyroid imaging reporting and data system[C-TIRADS]grading)and the laboratory marker(thyroid peroxidase antibody[TPOAb])were key predictive variables(all P<0.05).Among the four models,LightGBM performed the best:the AUC of the training set,the validation set and the independent external validation set was 0.978,0.918,and 0.889,respectively.Furthermore,it demonstrated significant clinical practicality in terms of calibration,clinical decision curve,and sensitivity.Conclusion LightGBM model can provide an efficient and stable auxiliary tool for early clinical differentiation of benign and malignant thyroid nodules in women.
作者
吴燕远
周薇
吴险
王健
WU Yanyuan;ZHOU Wei;WU Xian;WANG Jian(Key Laboratory of Clinical Laboratory Medicine of Guangxi Department of Education/Department of Clinical Laboratory,the First Affiliated Hospital of Guangxi Medical University,Nanning 530021,Guangxi,China;Department of Clinical Laboratory,the Eighth Affiliated Hospital of Guangxi Medical University—Guigang City People's Hospital,Guigang 537100,Guangxi,China)
出处
《右江医学》
2025年第8期708-718,共11页
Chinese Youjiang Medical Journal
基金
广西壮族自治区卫生健康委员会自筹经费科研课题(Z-R20241692,Z-R20231939)。
关键词
机器学习
女性甲状腺恶性结节
预测模型
C-TIRADS分级
LightGBM
machine learning
malignant thyroid nodules in women
predictive model
Chinese thyroid imaging reporting and data system(C-TIRADS)
LightGBM
