Objective：To explore the efficacy of a machine learning（ML）model utilizing multi-sequence MRI hand-crafted radiomics（HCR）and deep transfer learning（DTL）features to predict preoperative glioma grading. Methods：Image data were selected from 332 patients［258 cases of high-grade glioma（HGG），74 cases of low-grade glioma（LGG）］in the BraTS2019 dataset. A subset of 30 HGG and 8 LGG cases were randomly selected as test data set，while data from the remaining 294 cases were used as training and validation sets. HCR and DTL features of the lesion were extracted from T1，T2，T1c and Flair sequences，and the top 10 features were selected. Seven ML models were then constructed based on HCR features，DTL features，and a combination of both features of deep learning radiomics（DLR），to evaluate the efficiency of the models in predicting HGG and LGG. The SHAP method was subsequently employed to quantify and attribute the importance of features though analysis after selecting the best model. Results：The ML model constructed by DLR features based on combination of HCR and DTL exhibited the most superior predictive performance. Specifically， when features were filtered using the support vector machine-recursive feature elimination（SVM -RFE）method，the support vector machine（SVM）classifier model integrating T2，T1c and Flair sequences emerged as the best，achieving an area under receiver operating characteristic curve（AUC）of 0.996（95% CI：0.991-1.000）on the validation set，with a Youden index，accuracy，sensitivity and specificity of 0.920，0.976，0.988 and 0.932，respectively. On the test set，the model also displayed a great predictive capacity. SHAP analysis highlighted that the Flair sequence had the greatest feature contribution，followed by the T2 and T1c sequences. Both HCR and DTL teatures make significant contributions. Conclusion：The ML model，based on DLR features of multi-sequence MRI，is effective in glioma grade prediction. The optimal model is the SVM classifier using T2，T1c and Flair sequences after SVM-RFE selection.