A Comparative Analysis of the Usefulness of Survival Prediction Models for Patients with Glioblastoma in the Temozolomide Era: The Importance of Methylguanine Methyltransferase Promoter Methylation, Extent of Resection, and Subventricular Zone Location

OBJECTIVE: Several survival prediction models for patients with glioblastoma have been proposed, but none is widely used. This study aims to identify the predictors of overall survival (OS) and to conduct an independent comparative analysis of 5 prediction models. METHODS: Multi-institutional data from 159 patients with newly diagnosed glioblastoma who received adjuvant temozolomide concomitant chemoradiotherapy (CCRT) were collected. OS was assessed by Cox proportional hazards regression and adjusted for known prognostic factors. An independent CCRT patient cohort was used to externally validate the 1) RTOG (Radiation Therapy Oncology Group) recursive partitioning analysis (RPA) model, 2) Yang RPA model, and 3) Wee RPA model, Chai-chana model, and the RTOG nomogram model. The predictive accuracy for each model at 12-month survival was determined by concordance indices. Calibration plots were performed to ascertain model prediction precision. RESULTS: The median OS for patients who received CCRT was 19.0 months compared with 12.7 months for those who did not (P < 0.001). Independent predictors were: 1) subventricular zone II tumors (hazard ratio [HR], 1.6; 95% confidence interval [CI], 1.0-2.5); 2) methylguanine methyltransferase promoter methylation (HR, 0.36; 95% CI, 0.2-0.6); and 3) extent of resection of >85% (HR, 0.59; 95% CI, 0.4-0.9). For 12-month OS prediction, the RTOG nomogram model was superior to the RPA models with a c-index of 0.70. Calibration plots for 12-month survival showed that none of the models was precise, but the RTOG nomogram performed relatively better. CONCLUSIONS: The RTOG nomogram best predicted 12-month OS. Methylguanine methyltransferase promoter methylation status, subventricular zone tumor location, and volumetric extent of resection should be considered when constructing prediction models.