Using Machine Learning Models to Predict High-Grade Meningioma with Preoperative MR Imaging: A Meta-Analysis of Clinical Prediction Models

Background: Meningioma is the most common primary tumor of the central nervous system in adults. Higher-grade tumors significantly increase the risk of recurrence and decrease survival risk. Recent research has explored the use of magnetic resonance imaging (MRI) data with machine learning (ML) and deep learning (DL) algorithms to better predict WHO grade before surgery. This systematic review and meta-analysis assesses how accurately ML models, trained on preoperative MRI data, can differentiate between low-grade and high-grade meningiomas. Methods: MEDLINE, EMBASE, Web of Science, CENTRAL, Google Scholar, DBLP, IEEE Xplore, arXiv, and medRxiv databases were searched through May 2025 and studies involving ML models using preoperative MRI data to predict WHO grade were included. Data were extracted independently. A 2x2 table was constructed from data reported in the study or by sensitivity (Se)/specificity (Sp) and class counts, treating high-grade as positive. Data was synthesized using a bivariate random-effects model with forest plots of sensitivity and specificity, and an HSROC curve with a summary point. Study quality was assessed using the APPRAISE-AI tool. Results: Out of 1,492 records, 43 models across 27 studies ( n = 6,740) were included. The bivariate random-effects model estimated a summary sensitivity of 0.77 (95% CI, 0.72–0.82) and summary specificity of 0.86 (95% CI, 0.82–0.89). Between-study heterogeneity was moderate, with a generalized I 2 of 46.9% (Se I 2 = 42.9%; Sp I 2 = 51.6%). The HSROC showed a summary operating point near (Se 0.77, Sp 0.86) with wide prediction regions. Most studies (56%) achieved a high-quality rating, with 11% rated low and 30% moderate; one study (4%) reached very high quality. Recent studies (2020 onward) generally demonstrate higher quality than older ones. Conclusion: Machine learning models using preoperative MRI data show promising accuracy in distinguishing low- and high-grade meningiomas, with moderate sensitivity and high specificity. However, variability in study design and heterogeneity point to the need for standardized methods and external validation. Future research should focus on prospective, multicenter studies with transparent reporting to ensure ML tools can be reliably integrated into clinical decision-making. Publication History Article published online: 27 February 2026 © 2026. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany