Expanding the clinical utility of deep learning-based radiomics and machine learning for prognostic assessment in IDH-wildtype glioblastoma following maximal safe surgical resection

Dear Editor, We have carefully reviewed the recently published article in your journal entitled “Deep learning-based radiomics and machine learning for prognostic assessment in isocitrate dehydrogenase (IDH)-wildtype glioblastoma after maximal safe surgical resection: a multicenter study,” which integrates deep learning-based segmentation with radiomics for prognostic prediction in IDH-wildtype glioblastoma (GBM)[1]. This study marks a significant advancement in the field of precision oncology. However, several limitations must be addressed to enhance its clinical applicability. The article is compliant with the TITAN Guidelines 2025[2]. Limitations in sample size and external validation The study’s sample size, comprising 301 participants for training, 128 for internal validation, and 153 for external validation, may be inadequate given the 4227 radiomic features initially extracted. With a mortality rate of 60% in the training cohort, the effective sample size is approximately 180 events, resulting in a ratio of one event per 23 features, which falls short of the recommended 1:10 threshold[2]. Additionally, the external validation C-index of 0.76 represents a notable decrease from the internal validation C-index of 0.81, indicating potential overfitting issues. The heterogeneity observed between the training cohort (mean age 56.3 years) and the external validation cohort (mean age 60.8 years), particularly in terms of epidermal growth factor receptor amplification rates (54.8% vs. 38.6%), raises significant concerns regarding the generalizability of the model across diverse populations. Biological plausibility of selected features The biological plausibility of the selected features is inadequately addressed. Although the authors demonstrate the prognostic utility of features such as necrotic non-enhancing core and peritumoral edema, the mechanistic connections between these imaging biomarkers and the biology of GBM require further investigation. Recent studies have underscored the necessity of correlating radiomic signatures with underlying molecular pathways[3]. In the absence of such biological validation, the selected features may merely represent statistical associations rather than biologically meaningful predictors, thereby limiting their clinical credibility and acceptance within the neuro-oncology community. Model interpretability and clinical feasibility The absence of SHAP (SHapley Additive exPlanations) analysis or analogous interpretability tools hinders clinicians’ ability to comprehend the contribution of individual features to risk predictions[4]. This limitation is especially significant in the management of GBM, where treatment decisions necessitate transparent reasoning. The computational complexity involved in processing 4227 initial features through multiple algorithms may further obstruct real-time clinical implementation, particularly in resource-constrained settings. Comparative advantage over conventional methods necessitates quantification The study does not provide a direct comparison with established prognostic factors. Although the nomogram incorporates age and Karnofsky Performance Status (KPS) alongside radiomic scores, the incremental value over traditional clinical variables alone is not quantified. Previous studies have shown that simple clinical models, which include age, KPS, and molecular markers, can achieve C-indices ranging from 0.70 to 0.75[5]. Without a direct comparison or net reclassification improvement analysis, the added value of complex radiomic analysis remains uncertain, especially given the considerable computational resources required. We commend the authors for their innovative approach but encourage future studies to address these limitations through larger, more diverse cohorts, biological validation of features, incorporation of interpretability methods, and direct comparison with conventional prognostic models. These refinements will be essential for successful clinical translation of this promising technology.