Radiomic studies aim to predict survival in glioblastoma from MRI, but reported prognostic features vary widely across studies. This study asked whether the choice of statistical method itself explains that variation.

The authors pooled 705 IDH-wildtype glioblastoma patients from three public cohorts and tested 463 radiomic, anatomical, and clinical features against overall survival using three techniques: a log-rank test after median split, univariate Cox regression, and an XGBoost regressor interpreted with SHAP. The methods selected very different sets: 99, 72, and 20 features, respectively.

Only five features were chosen by all three: age, a peritumoural T2 texture feature, the number of connected peritumoural mask components, and two image-level intensity maxima. Where selections overlapped, effect directions agreed, but the overlap was limited—about half of the SHAP top 20 were also selected by either survival test.

The authors note that many selected features were PyRadiomics diagnostic descriptors not restricted to tumour voxels, and that analytical pipeline choices—ranking method, pre-filtering, descriptor handling—can substantially influence which features appear prognostic. This suggests that reported radiomic biomarkers may be as much a product of methodology as of biology.