AI–assisted survival prediction in glioma: Meta-analysis of prospective and controlled studies.

Abstract

e14000 Background: Accurate survival prediction in glioma is critical for individualized treatment planning, prognostic counseling, and clinical trial stratification. Artificial intelligence (AI)–based models, including machine learning and deep learning approaches, integrate radiomic, genomic, and clinical data beyond conventional prognostic tools. However, performance varies across studies, and comparative accuracy and generalizability remain uncertain. We conducted a meta-analysis of prospective and controlled studies evaluating AI-based overall survival (OS) prediction in adult glioma. Methods: PubMed, Embase, Cochrane Library, IEEE Xplore, and ClinicalTrials.gov were searched for prospective studies, controlled validation cohorts, or head-to-head comparisons published between January 2015 and December 2025. Eligible studies included adults (≥18 years) with histologically confirmed WHO grade II–IV glioma reporting OS prediction metrics (C-index, AUC, or Brier score) derived from MRI-based, radiogenomic, or multimodal AI models. Random-effects meta-analysis (DerSimonian–Laird) was performed. Heterogeneity was assessed using I², and risk of bias using PROBAST. Results: Twenty-two controlled studies including 9,314 patients met inclusion criteria. AI-based models demonstrated significantly higher OS prediction accuracy than standard prognostic tools. The pooled C-index for AI models was 0.81 (95% CI, 0.78–0.84; I² = 39%) compared with 0.69 (95% CI, 0.66–0.72) for conventional models, corresponding to a pooled mean improvement of 0.12 (p < 0.001). For 12-month OS prediction, AI models achieved a pooled AUC of 0.84 (95% CI, 0.80–0.87) versus 0.73 (95% CI, 0.70–0.76). Deep learning models showed higher accuracy (C-index 0.84) than traditional machine learning approaches (0.79). Multimodal AI models integrating imaging, genomic, and clinical data demonstrated the highest performance (C-index 0.87). External validation cohorts showed reduced accuracy (C-index 0.75), indicating limited generalizability. Acceptable calibration was reported in 64% of studies, while incomplete calibration reporting remained a major source of bias. Conclusions: AI-based survival prediction models significantly outperform conventional prognostic tools in glioma, particularly deep learning and multimodal approaches. Reduced performance in external validation highlights the need for larger, harmonized datasets and prospective trials to support clinical implementation. Pooled results of AI vs. standard prognostic models. Outcome AI Model (Pooled) Standard Model Effect Size 95% CI I² C-index (OS) 0.81 0.69 +0.12 0.09–0.15 39% AUC (12-mo OS) 0.84 0.73 +0.11 0.07–0.15 42% Deep Learning C-index 0.84 — — 0.81–0.87 33% Machine Learning C-index 0.79 — — 0.76–0.82 28% Multimodal AI C-index 0.87 — — 0.84–0.90 31% External Validation 0.75 — — 0.71–0.78 44%