Abstract Background Preoperative survival prediction in newly diagnosed glioblastoma (nGBM) remains challenging due to limited robustness and standardization across radiomic approaches. We aimed to validate a machine learning–based prognostic model using preoperative MR images and assess its generalizability. Methods Two independent cohorts were analyzed: the Kansai Molecular Diagnosis Network for CNS Tumors (KNBTG) and The Cancer Genome Atlas (TCGA). All cases with available preoperative MR images (T1WI, T2WI, and Gd-T1WI) were included. The KNBTG cohort was divided into a training dataset (TD, n = 137) and an internal test dataset (ITD, n = 141), while the TCGA cohort served as the external test dataset (ETD, n = 105). A total of 489 texture features were extracted. Overall survival (OS) was dichotomized at the median, and predictive modeling was performed using least absolute shrinkage and selection operator regularization. The trained model was validated on ITD and ETD. Results Radiomic high-risk status was associated with significantly shorter OS in both ITD and ETD (log-rank P. 05) and remained independently prognostic in multivariate Cox analysis. Time-dependent area under the receiver operating characteristic curves was consistently higher in models incorporating radiomic-based risk. Of the 13 selected features, “T2coreGLCMhomogeniety₃SD” was the only consistent predictor across cohorts and showed strong prognostic stratification, particularly between low- and high-risk groups (cutoff range: 0. 0145–0. 0180). Conclusions Radiomics-based modeling provides reproducible prognostic value in nGBM. The feature “T2coreGLCMhomogeniety₃SD” may serve as a reliable imaging biomarker for preoperative risk stratification and individualized treatment planning.
Umehara et al. (Thu,) studied this question.