谌典,周畅,张奥懿,聂淑婷,邵袁缘,鲜锋,胡文姝,李心怡.基于临床、超声特征及影像组学构建机器学习模型预测慢性肾脏病患者肾功能损伤程度[J].中国医学影像技术,2024,40(4):575~579
基于临床、超声特征及影像组学构建机器学习模型预测慢性肾脏病患者肾功能损伤程度
Machine learning model based on clinical, ultrasonic features and radiomics for predicting renal function damage degree in patients with chronic kidney disease
投稿时间:2023-12-19  修订日期:2024-01-28
DOI:10.13929/j.issn.1003-3289.2024.04.021
中文关键词:  肾功能不全,慢性  超声检查  影像组学  机器学习
英文关键词:renal insufficiency, chronic  ultrasonography  radiomics  machine learning
基金项目:
作者单位E-mail
谌典 三峡大学第一临床医学院 宜昌市中心人民医院超声科, 湖北 宜昌 443000  
周畅 三峡大学第一临床医学院 宜昌市中心人民医院超声科, 湖北 宜昌 443000 zhouch2004@126.com 
张奥懿 三峡大学第一临床医学院 宜昌市中心人民医院超声科, 湖北 宜昌 443000  
聂淑婷 三峡大学第一临床医学院 宜昌市中心人民医院超声科, 湖北 宜昌 443000  
邵袁缘 三峡大学第一临床医学院 宜昌市中心人民医院超声科, 湖北 宜昌 443000  
鲜锋 三峡大学第一临床医学院 宜昌市中心人民医院超声科, 湖北 宜昌 443000  
胡文姝 三峡大学第一临床医学院 宜昌市中心人民医院超声科, 湖北 宜昌 443000  
李心怡 三峡大学第一临床医学院 宜昌市中心人民医院超声科, 湖北 宜昌 443000  
摘要点击次数: 306
全文下载次数: 338
中文摘要:
      目的 观察基于临床、超声特征及影像组学构建机器学习(ML)模型预测慢性肾脏病(CKD)患者肾功能损伤程度的价值。方法 回顾性分析199例CKD患者资料,以9 ∶ 1比例将其分为训练集(n=179)及验证集(n=20),根据估算肾小球滤过率(eGFR)划分轻中度或重度肾功能损伤。采用多因素logistic回归分析训练集临床及超声特征,筛选CKD患者肾功能损伤程度的独立预测因素,分别基于支持向量机(SVM)、极致梯度提升(XGBoost)及逻辑回归(LR)算法构建临床-超声模型、影像组学模型及联合模型;绘制受试者工作特征(ROC)曲线,计算曲线下面积(AUC),评估各模型预测CKD患者肾功能损伤程度的效能。结果 肾脏长径为CKD患者肾功能损伤的独立预测因素(P<0.05)。以不同算法所获模型中,以SVM算法所获临床-超声模型、影像组学模型及联合模型预测肾功能损伤的效能最高;训练集中,以SVM算法所获临床-超声模型的敏感度、特异度、准确率及AUC分别为81.93%,62.50%,71.51%及0.722,影像组学模型分别为89.16%,70.83%,79.33%及0.800,联合模型分别为84.34%,80.21%,82.12%及0.822;验证集中,以SVM算法所获临床-超声模型的敏感度、特异度、准确率及AUC分别为75.00%、66.67%、70.00%及0.708,影像组学模型分别为75.00%、58.33%、65.00%及0.667,联合模型分别为87.50%、75.00%、80.00%及0.812。结论 基于超声特征联合影像组学构建的ML模型可有效预测CKD患者肾功能损伤程度;利用SVM算法获得的联合模型具有最佳效能。
英文摘要:
      Objective To observe the value of machine learning (ML) models based on clinical, ultrasonic features and radiomics for predicting renal function damage degree in patients with chronic kidney disease (CKD). Methods Data of 199 CKD patients were retrospectively analyzed. The patients were randomly divided into training set (n=179) and validation set (n=20) at the ratio of 9 ∶ 1, and further classified as mild-moderate or severe renal function damage according to estimated glomerular filtration rate (eGFR). Multivariate logistic analysis was used to analyze clinical and ultrasonic features, so as to screen the independent predictors of renal function damage degree of CKD patients. Then clinical-ultrasonic model, radiomics model and combined model were constructed using support vector machine (SVM), extreme gradient boosting (XGBoost) and logistic regression (LR), respectively. Receiver operating characteristic (ROC) curves were drawn, the area under the curves (AUC) were calculated to evaluate the efficacy of each model for predicting renal function damage degree of CKD patients. Results Renal length was an independent predictive factors for renal function damage degree of CKD patients (P<0.05). Among models obtained with different algorithms, modelclinical-ultrasound, modelradiomics and modelcombination obtained with SVM had the highest prediction efficacy, in training set, the sensitivity, specificity, accuracy and AUC of SVM modelclinical-ultrasound was 81.93%, 62.50%, 71.51% and 0.722, of SVM modelradiomics was 89.16%, 70.83%, 79.33% and 0.800, of SVM modelcombination was 84.34%, 80.21%, 82.12% and 0.822, respectively, in validation set, the sensitivity, specificity, accuracy and AUC of SVM modelclinical-ultrasound was 75.00%, 66.67%, 70.00% and 0.708, of SVM modelradiomics was 75.00%, 58.33%, 65.00% and 0.667, of SVM modelcombination was 87.50%, 75.00%, 80.00% and 0.812, respectively. Conclusion ML models based on ultrasonic features and radiomics could be used to predict renal function damage degree in patients with CKD, and SVM modelcombination had the best efficacy.
查看全文  查看/发表评论  下载PDF阅读器