Purpose: Identifying radiomics features that help predict whether glioblastoma patients are prone to developing epilepsy may contribute to an improvement of preventive treatment and a better understanding of the underlying pathophysiology.

Materials and methods: In this retrospective study, 3-T MRI data of 451 pretreatment glioblastoma patients (mean age: 61.2 ± 11.8 years; 268 men, 183 women) were analyzed. Three hundred thirty-six patients reported no epilepsy, while 115 patients were diagnosed with symptomatic epilepsy. A total of 1,546 radiomics features were extracted from contrast-enhancing tumor, peritumoral regions, and normal-appearing white matter as regions of interest using PyRadiomics. The dataset was initially split into a training (70%) and a validation (30%) cohort. The training cohort was used for feature selection with ElasticNet and model optimization. Various machine learning models, including logistic regression (LR), were used to predict epilepsy status. The models' performances were evaluated with the validation cohort, and the area under the curve of the receiver operating characteristics (AUC) was used as a measure. For identifying relevant features, permutation feature importance was applied.

Results: The performance of LR using radiomics features from only a single ROI in the validation cohort was AUC = 0.83 (95% CI: 0.76-0.91) and AUC = 0.77 (95% CI: 0.69-0.85) for the peritumoral and white matter regions, respectively. The most important features in peritumoral regions were shape features, while for the white matter region, higher-order features from FLAIR were most relevant.

Conclusion: Radiomics features from peritumoral and normal-appearing white matter can be associated with epilepsy status at diagnosis, suggesting an important role of these regions for the development of epilepsy in glioblastoma patients.