Objective:To explore the application value of pancreatic CT-based radiomics in predicting insulin resistance(IR)in people with impaired glucose tolerance.Methods:A total of 381 patients initially diagnosed with impaired glucose tolerance were retrospectively collected and were divided into two groups based on homeostasis model assessment of IR(HOMA-IR),a high-IR group(191 cases)and a low-IR group(190 cases).And all patients were randomly divided into the training cohort and the validation cohort at a ratio of 8∶2.Pancreatic ROIs were sketched and radiomics features were extracted,and the optimal features were selected after dimensionality reduction and screening.Eight machine learning models were constructed,and four machine learning methods(SVM,MLP,RF and AdaBoost)were selected to construct the diagnostic prediction model.ROC curve was used to evaluate the prediction performance of each radiomics model.Results:A total of 1 834 features were extracted and 189 features were screened by Pearson's correlation analysis.The dimensionality was reduced to 23 major radiomics features by LASSO method and 5-fold cross-validation.The AUCs of the four constructed prediction models based on SVM,MLP,RF and AdaBoost for the validation cohort were 0.723,0.731,0.807 and 0.681,respectively,with the better prediction efficiency of RF.Conclusion:The RF model based on pancreatic CT radiomics features has a better potential prediction for the IR level in people with impaired glucose tolerance.
国家科技期刊平台
NSTL
万方数据
