Interpreting Multimodal Data Fusion for Glioblastoma Prognosis: A Systematic Literature Review of Explainable AI Applications

Abstract

A representative clinical scenario involves a 57-year-old patient whose brain MRI reveals an infiltrative mass with central necrosis, with biopsy confirming glioblastoma (GBM, WHO Grade 4). Despite maximal resection, concurrent temozolomide, and radiotherapy, median survival remains 14–16 months, a figure that has remained largely unchanged since 2005. This persistent survival plateau highlights a deeper diagnostic limitation: conventional prognostic tools account for age and surgical extent but overlook molecular determinants such as IDH mutation status, MGMT promoter methylation, and transcriptomic subtype, which critically influence patient outcomes. Multimodal data fusion directly addresses this gap by integrating multi-parametric MRI (mpMRI), PET imaging, next-generation genomics, and clinical variables (age, Karnofsky Performance Status [KPS]) into a unified predictive framework. Deep learning architectures leveraging these multimodal inputs have achieved Concordance Indices (C-indices) up to 0.80, compared to 0.65 for unimodal approaches. However, interpretability remains a significant barrier to clinical adoption: models such as Convolutional Neural Networks (CNNs) or Vision Transformers (ViTs) that surpass clinician performance on test sets cannot inherently explain their risk estimates. Without interpretability, these models cannot be validated against biological mechanisms, reconciled with clinical judgment, or approved by regulatory authorities. Explainable AI (XAI) methods, including SHAP (Shapley Additive exPlanations), Grad-CAM, and cross-modal attention mechanisms, translate opaque model outputs into biologically interpretable rationales. This PRISMA-compliant systematic review synthesizes 22 peer-reviewed studies (2023–2026), finding that hybrid transformer-based fusion dominates the field (45% of studies, C-index 0.71–0.80), with SHAP (55%) and attention maps (32%) leading in XAI adoption. Notably, XAI-integrated models have advanced beyond replicating established biomarker associations: cross-modal attention analyses have uncovered a previously undocumented interaction between FLAIR-hyperintense peritumoral oedema and EGFR amplification that predicts early recurrence, demonstrating that interpretability tools facilitate biological discovery as well as accountability.