Machine learning combined with computed tomography radiomics in predicting vertebral fragility fractures in patients with type 2 diabetes mellitus

Abstract:

Objective To investigate the accuracy of a model established based on machine learning and computed tomography（CT） radiomics features in predicting vertebral fragility fractures in patients with type 2 diabetes mellitus（T2DM）.Methods A retrospective analysis was performed for the CT images and clinical data of 140 patients，among whom there were 70 T2DM patients with newly diagnosed vertebral fragility fractures and 70 patients in the control group. The previous CT images and clinical data of 18 patients（16 T2DM patients with vertebral fragility fractures and 2 patients in the control group） were collected as an external validation set. The optimal features were screened by the univariate analysis，the Pearson correlation analysis，minimum redundancy maximum relevance algorithm，the binary logistic regression analysis，and the least absolute shrinkage and selection operator regression model，and then a predictive model was constructed by support vector machine，multi-layer perceptron，and eXtreme gradient boosting（XGBoost） classifiers. The area under the ROC curve（AUC） was used to evaluate the predictive performance of the model.Results A total of 1 037 radiomics features were extracted from the CT images of each patient and were then simplified into 14 radiomics features. Among the 17 clinical features，sex，age，and body mass index were independent factors for predicting outcome. XGBoost classifier showed the best performance，and the XGBoost model showed an AUC of 1.000，0.929，and 1.000，respectively，in the training set and an AUC of 0.954，0.862，and 0.969，respectively，in the test set.Conclusion The XGBoost model based on clinical and radiomics features can be used as a noninvasive tool for predicting vertebral fragility fractures in T2DM patients.