目的 分析影像组学特征联合机器学习模型对慢性阻塞性肺疾病急性加重期(AECOPD)患者预后的预测价值.方法 回顾性分析2016年7月—2021年2月于华北理工大学附属医院诊断为AECOPD的121例患者,收集其一般资料,采用Pyradiomics软件包提取胸部CT横断位影像组学特征.根据随访结果将患者分为预后不良组(n=59)和预后良好组(n=62).采用Lasso回归模型对影像组学特征进行筛选降维,并计算radscore;按7︰3的比例将患者分为训练集(n=85)与测试集(n=36),分别采用LR算法、SVM算法、DT算法构建影像组学预测模型,采用ROC曲线、混淆矩阵及De-long检验,评估三种模型对AECOPD预后的预测效能并比较模型之间的差异.结果 共提取出2 381个影像组学特征.Lasso回归模型结果显示,模型拟合效果最好时共包含11个影像组学特征.将这11个影像组学特征与相应的加权系数乘积的线性组合作为rad_score.分析结果显示,LR算法、SVM算法、DT算法三种模型预测AECOPD预后的效能不同,在训练集中其AUC值分别为0.858、0.774和0.767,测试集中分别为0.848、0.748、0.774;其中LR模型的AUC值均最高,约登指数显示其预测效能最优.Delong检验结果显示在训练集中和测试集中,LR模型的AUC值均高于SVM模型(0.858 vs 0.774,P=0.022;0.848 vs 0.748,P=0.011);训练集与测试集中LR模型与DT模型、SVM模型与DT模型的AUC比较差异均无统计学意义(P＞0.05).综合AUC值、约登指数与Delong检验结果,LR模型在此次研究中表现最优.结论 由LR算法、SVM算法、DT算法三种机器学习算法构建的影像组学预测模型对AECOPD预后有较好的预测价值,其中LR模型最优.
Analysis of the predictive value of CT radiomics features combined with machine learning for prognosis of AECOPD
Objective To analyze the predictive value of radiomics features combined with ma-chine learning models for the prognosis of patients with acute exacerbations of chronic obstructive pul-monary disease(AECOPD).Methods A retrospective analysis was conducted on 121 patients diag-nosed with AECOPD at the Affiliated Hospital of North China University of Science and Technology from July 2016 to February 2021.General information was collected,and radiomics features were ex-tracted from transverse chest CT images using the Pyradiomics software package.Patients were divid-ed into a poor prognosis group(n=59)and a good prognosis group(n=62)based on follow-up re-sults.The Lasso regression model was used to screen and reduce the dimensionality of radiomics fea-tures,and the rad_score was calculated.AECOPD patients were divided into a training set(n=85)and a test set(n=36)at a ratio of 7︰3.Radiomics prediction models were constructed using LR,SVM,and DT algorithms,respectively.ROC curves,confusion matrices,and Delong tests were used to evaluate the predictive performance of the three models for AECOPD prognosis and compare the differences between the models.Results A total of 2 381 radiomics features were extracted.The Lasso regression model results showed that the best model fit included 11 radiomics features.The linear com-bination of these 11 radiomics features and their corresponding weighted coefficients was used as the rad_score.The analysis results showed that the predictive performance of the three radiomics models constructed using LR,SVM,and DT algorithms for AECOPD prognosis varied.In the training set,the AUC values were 0.858,0.774,and 0.767,respectively,and in the test set,they were 0.848,0.748,and 0.774,respectively.Among them,the LR model had the highest AUC value,and the Youden index indicated that it had the best predictive performance.The Delong test results showed that in both the training and test sets,the AUC value of the LR model was higher than that of the SVM model(0.858 vs.0.774,P=0.022;0.848 vs.0.748,P=0.011).The differences in AUC be-tween the LR model and the DT model,as well as between the SVM model and the DT model,in both the training and test sets were not statistically significant(P＞0.05).Considering the AUC values,Youden index,and Delong test results,the LR model performed best in this study.Conclusion Ra-diomics prediction models constructed using LR,SVM,and DT machine learning algorithms have good predictive value for AECOPD prognosis,with the LR model performing the best.
RadiomicsAcute exacerbations of chronic obstructive pulmonary diseaseMa-chine learning classifiers
万方数据
维普
