Diagnostic accuracy of machine learning models in glioma classification: A meta-analysis.

Abstract

2079 Background: Machine learning (ML) is promising in IDH-based glioma classification using magnetic resonance imaging (MRI), but variability in methods and algorithms necessitates a comprehensive evaluation. This meta-analysis assesses the pooled diagnostic performance of ML-based approaches. Methods: A literature search was conducted in January 2025 across PubMed, MEDLINE, and Cochrane. MICCAI, RSNA, and SNO meeting abstracts were additionally reviewed. Eligible studies evaluating ML models for IDH-based glioma classification using MRI were included. Data were pooled using a random-effects model, analyzing sensitivity, specificity, heterogeneity, and publication bias via Egger’s test and funnel plots. Leave-one-out analysis was conducted. Results: A total of 5982 cases were analyzed. Gliomas were classified as WHO Grades II (11.5%), III (23.1%), and IV (65.4%). Histopathology-based reference standards, including genetic and molecular testing, were used in 73.1% of studies, while immunohistochemistry, pathology, biopsy-proven markers, and immunohistopathologic diagnosis were each used in 3.8–7.7%. Deep learning models, including CNNs and ResNet, were the most used classifiers (30.8%), followed by Support Vector Machines (26.9%). Ensemble methods, such as Random Forest accounted for 19.2%, regression-based approaches (LASSO, logistic regression) for 15.3%, and other techniques like multilayer perceptron and AdaBoost for 7.7%. This meta-analysis included 25 studies for sensitivity and 26 for specificity, using a random-effects model with DerSimonian-Laird estimation. Pooled sensitivity was 83.0% (95% CI: 79.5–86.5%) and specificity was 78.6% (95% CI: 73.7-83.4%), both statistically significant (p < 0.0001). Substantial heterogeneity was found (I² = 100% for both), with Cochran’s Q values of 1.9e+06 for sensitivity and 4.2e+06 for specificity (p < 0.001). Leave-one-out analysis showed minimal variation in pooled estimates (sensitivity: 82.6–83.8%, specificity: 77.6–79.4%). Egger’s test revealed significant small-study effects (p = 0.0001 for both), suggesting potential publication bias. Conclusions: ML models demonstrated moderate diagnostic performance in IDH-based glioma classification, achieving a sensitivity of 83.1% and specificity of 78.6%. However, substantial heterogeneity and potential biases pose significant challenges to their clinical implementation. To enhance the reliability and broader applicability of ML models in IDH-based glioma diagnosis, standardization of imaging protocols and external validation are imperative. Meta-analytical findings. Metric Sensitivity (%) Specificity (%) Pooled Estimate 83.0 (79.5–86.6) 78.6 (73.7–83.4) Z-Value 46.4 31.76 P-Value <0.0001 <0.0001 I² (%) 100 100 T² 80.0 159.0 Cochran's Q 1.9e+06 (df=24, p<0.001) 4.2e+06 (df=25, p<0.001) Leave-One-Out Range 82.6–83.8 77.6–79.4 Egger’s Test (P-value) 0.001 0.001