Diffuse gliomas represent a highly heterogeneous group of central nervous system neoplasms characterized by a spectrum of aggressive malignant phenotypes, with glioblastoma in particular being associated with a median survival of less than fifteen months. The 2021 WHO Classification of Tumors of the Central Nervous System (CNS5) established a paradigm shift toward integrated diagnostics, necessitating the combination of histopathological assessment with definitive molecular markers, such as IDH mutation status, MGMT promoter methylation, and 1p/19q codeletion. While unimodal artificial intelligence (AI) models trained on magnetic resonance imaging (MRI) have demonstrated efficacy in structural segmentation, they frequently fail to reliably resolve these critical molecular characteristics required for targeted therapies and surgical planning. Although recent technical advancements have introduced complex multimodal data fusion strategies—leveraging Transformers, Graph Neural Networks (GNNs), and hybrid architectures—to bridge the gap between imaging, clinical, and molecular profiling, current systematic reviews often aggregate broad diagnostic outcomes without providing a granular taxonomy of fusion methodologies or evaluating specific efficacy across diverse datasets. Furthermore, pervasive "black-box" limitations and a lack of cross-vendor validation remain significant barriers to clinical translation, yet current literature offers no comparative quantitative evidence on which architectures best mitigate these issues. The primary objective of this research is to systematically evaluate, categorize, and quantitatively compare the diagnostic performance of three distinct multimodal data fusion strategies: early (feature-level), late (decision-level), and joint/hybrid architectures for AI-based glioma classification and grading. By establishing a rigorous taxonomy of these fusion paradigms, this review aims to identify the optimal integration strategies for disparate clinical data streams, including MRI-radiomics, histopathology, genomics, and clinical demographics. To establish the incremental clinical value and justify the computational complexity of these multimodal architectures, a co-primary objective is to evaluate their diagnostic superiority or non-inferiority against unimodal AI baselines and standard clinical reference tests. Secondary objectives include assessing predictive accuracy across key 2021 WHO CNS5 integrated diagnostic markers, auditing technical pipelines for spatial co-registration and class imbalance handling, and identifying specific architectures that best mitigate "black-box" limitations to maximize clinical interpretability. The study follows a rigorous methodological framework anchored to the PRISMA 2020 statement, implementing a pre-specified statistical framework designed to harmonize disparate performance metrics historically characterized by pervasive heterogeneity. The expected outcomes of this investigation include the delivery of the first structured, quantitative comparison of fusion paradigms in neuro-oncology, identifying the most robust architectural strategies for diagnostic classification. The meta-analysis will utilize a random-effects model using the Restricted Maximum Likelihood (REML) estimator on the logit scale to synthesize pooled estimates for primary outcomes, specifically the discriminative performance in distinguishing low-grade from high-grade gliomas. A critical expected outcome is the establishment of a methodological benchmark for mitigating statistical heterogeneity in the evidence synthesis of medical AI. Additionally, the research will pioneer a Preprint-to-Publication Outcome Audit to detect "outcome switching" and implement an AI-Adapted GRADE framework to adjudicate the certainty of cumulative evidence based on a Fusion Complexity Gradient. Ultimately, this work is intended to provide a comprehensive roadmap for regulatory-grade validation, facilitating the clinical translation of advanced multimodal AI systems for non-invasive precision neuro-oncology.
Amir Mohammad Khajeh Aminian (Tue,) studied this question.