A two-stage model combining clinical and delta-radiomics features for prognosis in high-grade gliomas after chemoradiotherapy: a retrospective study.

OBJECTIVE: This study aims to evaluate the efficacy of integrating pre-radiotherapy and mid-radiotherapy radiomics, temporal changes in tumor radiomics (delta radiomics) during chemoradiotherapy, and clinical variables into prognostic models for individualized overall survival (OS) prediction in patients with high-grade glioma (HGG). METHODS: A total of 143 patients with HGG who underwent standard chemoradiotherapy were retrospectively enrolled from two centers. Patients were divided into a training cohort (n = 116), in which 5-fold cross-validation was performed, and an external test cohort (n = 27). Radiomics features were extracted from pre-radiotherapy and mid-radiotherapy MRI scans, and delta radiomic features were calculated as the differences between the two timepoints. Feature selection was performed using elastic-net-based stability selection, and radiomics scores (Rad-scores) were generated for each selected feature set using Cox models. Rad-scores were subsequently combined with clinical variables in a random survival forest (RSF) model for OS prediction. Kaplan-Meier survival analysis was performed to evaluate prognostic stratification. RESULTS: The combined model integrating clinical variables and delta radiomics achieved a C-index of 0.818 and 0.761 in the validation and external test cohorts, respectively, outperforming models based solely on radiomics or clinical features. In the external test cohort, time-dependent discrimination yielded 12-, 18-, 24- and 30-month OS AUCs of 0.842, 0.856, 0.796, and 0.943, and Kaplan-Meier analysis showed significant risk stratification (p = 0.002). CONCLUSION: The findings of this retrospective study suggest that integrating multi-timepoint radiomics, delta radiomics features, and clinical variables may facilitate OS prediction in patients with HGG following standard chemoradiotherapy. While promising, the exploratory findings require further validation in larger prospective cohorts, given the limited size of the external validation cohort and the absence of MGMT promoter methylation status from the main model.