Prediction of preoperative tumor-related epilepsy using XGBoost radiomics models with 4 MRI sequences

<i>Introduction</i>. Tumor-related epilepsy is a prevalent condition in patients with gliomas. Accurate prediction of epilepsy is crucial for early treatment. This study aimed to evaluate the novel application of the eXtreme Gradient Boost (XGBoost) machine learning (ML) algorithm into a radiomics model predicting preoperative tumor-related epilepsy (PTRE). Its performance was compared with 4 conventional ML algorithms, including the least absolute shrinkage and selection operator (LASSO), elastic net, random forest, and support vector machine.<i>Methods.</i>This study used four magnetic resonance imaging (MRI) images consisting of four sequences (T1-weighted [T1W], T1-weighted contrast [T1WC], T2-weighted [T2W], and T2-weighted fluid-attenuated inversion recovery [T2W FLAIR]) acquired from 74 glioma patients, 30 with PTRE and 44 without PTRE. 394 radiomics features were extracted from the MRI scans using<i>Pyradiomics</i>, alongside 12 clinical features from the medical records. The ML algorithms were mixed and matched to create 20 radiomics models with two stages for: (1) feature selection and (2) prediction of PTRE. Nested cross-validation was used to tune the algorithms and select the stable features.<i>Results.</i>The XGBoost radiomics model demonstrated the second-highest balanced accuracy and F1-score of 0.81 ± 0.01 and 0.80 ± 0.01 respectively. It also achieved the highest recall of 0.81 ± 0.02. It used mostly textural radiomics features from the T1W, T2W and T2W FLAIR sequences to make the predictions.<i>Conclusion.</i>This study demonstrates that XGBoost is a viable alternative to conventional ML algorithms for developing a radiomics model to predict PTRE, as the model produced from XGBoost had among the highest metrics. XGBoost selected features with a higher predictive value than other models. The features selected by XGBoost were more stable, which is a useful property for radiomics analysis. Features selected from multiple MRI sequences were important in the model's decision.