Predicting Recurrent Patterns After Radiotherapy Plus Concomitant and Adjuvant Temozolomide in Patients With High-Grade Gliomas Incorporating Multimodal Magnetic Resonance Imaging and Molecular Biomarkers

Abstract

This study aimed to establish a multivariable logistic regression model based on multimodal magnetic resonance (MR) imaging and molecular biomarkers to provide individualized prediction of recurrent patterns (RPs) in patients with high-grade glioma (HGG) after temozolomide-based chemoradiotherapy.We retrospectively analyzed 152 patients with recurrent HGG who had contemporaneous diffusion tensor imaging (DTI) and MR spectroscopy (MRS) at initial diagnose. Patients were randomly divided into training (n = 102) and internal validation cohorts (n = 50). Local and distant recurrent were assessed, with the pattern of recurrence of individual lesions defined relative to the 95% isodose line. DTI parameters of the tumor, including fractional anisotropy (FA) and apparent diffusion coefficient were calculated. N-acetylaspartate and choline (Cho) values of the tumor were measured from MRS and normalized as ratios to creatine (Cr). Clinical data collected were grade, location and molecular status. Logistic regression analysis was used to identify risk factors for building the prediction model in training cohorts. The results were internally validated by assessment of discrimination and calibration using the validation cohorts.For all the patients, local recurrences occurred in 105 (69.08%) patients, and distant recurrent developed in 47 (30.92%) patients. Multivariable logistic analysis indicated that grade (OR: 4.585; 95% CI: 1.082-19.436), subventricular zone (SVZ) contact (OR: 6.893; 95% CI: 1.325-35.868), FA value of tumor (OR: 6.989; 95% CI: 1.310-37.271), Cho/Cr ratio (OR: 14.327; 95% CI: 2.697-76.112) and O6-methylguanine-DNA methyltransferase (MGMT) methylation status (OR: 5.467; 95% CI: 1.453-20.577) were independent predictors of RPs. A prediction model was developed as following: Logit (P) = -7.038 + 1.523XGrade + 1.931XSVZ + 1.944XFA + 2.662XCho/Cr + 1.699XMGMT. Our model showed superior discriminatory power in the training cohort (AUC = 0.897, P < 0.001). Calibration curves indicated favorable consistency between the model prediction and the actual outcomes (Hosmer-Lemeshow test = 0.853). Moreover, the decision curve analyses exhibited satisfactory clinical utility. The favorable performance of this prediction model was confirmed in the internal validation cohorts.Multimodal MR imaging can be integrated with molecular biomarkers to provide better RPs prediction for HGG patients. The developed model could help the clinicians to make decision-making in patients with HGG undergoing (chemo) radiotherapy.S. Nie: None. Y. Wang: None. X. Ding: None. Z. Zhou: None. Y. Guo: None. M. Hu: None.