基于18F-FDG PET/CT的机器学习模型对风湿性多肌痛的诊断价值
Diagnostic value of machine learning model based on 18F-FDG PET/CT for polymyalgia rheumatica
孙苏文冬 1邵晓梁 2蒋婉岚 1张璐 1徐婷 1吴敏 1王跃涛2
作者信息
- 1. 苏州大学附属第三医院、常州市第一人民医院免疫风湿科,常州 213003
- 2. 苏州大学附属第三医院、常州市第一人民医院核医学科,常州 213003
- 折叠
摘要
目的 探讨基于18F-FDG PET/CT的机器学习模型对风湿性多肌痛(PMR)的诊断价值.方法 回顾性分析2014年11月至2022年12月间入住常州市第一人民医院免疫风湿科,疑似PMR并行18F-FDG PET/CT检查的177例患者[男119例、女58例,年龄67.0(61.0,72.0)岁]资料.将患者按照7∶3随机等比抽样分为训练集和验证集.利用分类和回归树(CART)、最小绝对收缩和选择算子(LASSO)算法和logistic回归3种机器学习模型对PET/CT影像学特征进行学习.通过ROC曲线分析评估各模型的诊断效能,采用Delong检验比较不同AUC的差异.结果 PMR患者78例(44.1％,78/177),非 PMR 患者 99 例(55.9％,99/177);训练集 124 例,验证集 53 例.Logistic 回归模型(训练集:AUC=0.961;验证集:AUC=0.930)在诊断PMR方面优于CART(训练集:AUC=0.902,z=2.96,P=0.003;验证集:AUC=0.844,z=2.46,P=0.014),与 LASSO 算法诊断效能相似(训练集:AUC=0.957,z=0.95,P=0.340;验证集:AUC=0.930,z=0.00,P=1.000),但其评估部位较少.简化后的PMR-Logit评分在总体人群中的AUC为0.951,诊断PMR的灵敏度为89.74％(70/78),特异性为90.91％(90/99).结论 基于18F-FDG PET/CT影像学特征的机器学习模型有望成为一种有效诊断PMR的工具.
Abstract
Objective To investigate the diagnostic value of machine learning model based on18F-FDG PET/CT for polymyalgia rheumatica(PMR).Methods From November 2014 to December 2022,177 patients(119 males,58 females;age:67.0(61.0,72.0)years)admitted to the Department of Rheu-matology and Immunology,the First People's Hospital of Changzhou,with suspected PMR and undergoing 18F-FDG PET/CT examination were retrospectively analyzed.Patients were randomly divided into training set and validation set at the ratio of 7∶3.Three machine learning models,including classification and regression tree(CART),the least absolute shrinkage and selection operator(LASSO)algorithm,and logistic regres-sion,were established based on the PET/CT imaging features to aid in the diagnosis of PMR.The diagnostic efficacy of each model was evaluated by ROC curve analysis and differences among AUCs were analyzed by Delong test.Results There were 78(44.1％,78/177)PMR patients and 99(55.9％,99/177)non-PMR pa-tients,and 124 patients in the training set and 53 patients in the validation set.The logistic regression model(training set:AUC=0.961;validation set:AUC=0.930)was superior to the CART(training set:AUC=0.902,z=2.96,P=0.003;validation set:AUC=0.844,z=2.46,P=0.014)in diagnosing PMR,and was similar to LASSO algorithm(training set:AUC=0.957,z=0.95,P=0.340;validation set:AUC=0.930,z=0.00,P=1.000),but with fewer sites evaluated.The simplified PMR-Logit score had the AUC of 0.951 in the overall population,with the sensitivity of 89.74％(70/78)and the specificity of 90.91％(90/99).Conclu-sion Machine learning models based on 18F-FDG PET/CT imaging features are expected to be an effective diagnostic tool for PMR.
关键词
风湿性多肌痛/影像组学/机器学习/正电子发射断层显像术/体层摄影术,X线计算机/氟脱氧葡萄糖F18Key words
Polymyalgia rheumatica/Radiomics/Machine learning/Positron-emission tomo-graphy/Tomography,X-ray computed/Fluorodeoxyglucose F18引用本文复制引用
基金项目
常州市"十四五"卫生健康高层次人才培养工程领军人才(2022260)
出版年
2024
NETL
NSTL
万方数据