DSAI-08 PREDICTION MODELS FOR RADIATION NECROSIS AFTER STEREOTACTIC RADIOSURGERY FOR BRAIN METASTASES: A SYSTEMATIC REVIEW AND META-ANALYSIS
Ruchit Jain, Fatma Nihan Akkoc Mustafayev, Namita Ruhela, Zouina Sarfraz, Khalid Ahmad Qidwai, Manmeet Singh AhluwaliaAbstract
Background
Radiation necrosis (RN) is a clinically significant toxicity following stereotactic radiosurgery (SRS) for brain metastases, with increasing relevance in the era of multimodal therapy. Although multiple prediction models have been proposed, their performance has not been systematically studied. We performed a meta-analysis to evaluate the performance of RN prediction models.
Methods
A systematic search of PubMed, Embase, and Scopus and Web of Science was conducted to identify studies evaluating prediction models for RN following SRS. Studies reporting model performance using area under the receiver operating characteristic curve (AUC) or C-statistics were included. One model per study was extracted, prioritizing the best-performing validated model. AUCs and C-statistics were pooled using a random-effects model, with standard errors derived from reported confidence intervals.
Results
Five studies comprising 719 patients and 1,165 lesions were included. Prediction models included NTCP-based, radiomics-based, and deep learning approaches. RN rates ranged from 8% to 27% across studies. The pooled AUC was 0.76 (95% CI 0.61–0.92), with substantial heterogeneity (I² = 84.8%). AI-based models, particularly ensemble and deep learning approaches, demonstrated higher discrimination (AUC up to 0.91) compared with traditional dosimetric models.
Conclusions
Prediction models for RN after SRS show moderate performance overall, with higher accuracy achieved by AI-based approaches. However, substantial heterogeneity and limited external validation constrain clinical applicability. Standardized, prospectively validated models are needed before routine clinical implementation.