Evaluating Single and Multi-Omics Based Explainable Artificial Intelligence (MOXAI) for Molecular Subclass Classification of Adult-Type Diffuse Gliomas

Multimodal AI improves diagnostic accuracy: The combined analysis of DNA methylation and copy number data achieved 92.98% accuracy in classifying brain tumor subtypes, outperforming models that analyzed these data separately—demonstrating that integrating complementary biological information produces more reliable AI-driven diagnoses.

Explainability reveals which genetic features drive decisions: The researchers developed methods (CAMs and Grad-CAM) that identify the specific genetic locations and genes responsible for each classification decision, addressing a critical gap in medical AI where doctors need to understand why the system made a particular diagnosis rather than just accepting a "black box" prediction.

Risk stratification within tumor categories becomes possible: Unlike traditional approaches, MOXAI can assess risk levels within a single tumor subclass, potentially enabling more personalized treatment recommendations and better prognostic information for patients with the same tumor type.

Model generalization matters for clinical deployment: The DINOv2 model achieved 94.25% accuracy on independent validation data with high confidence (>0.9), suggesting the system can reliably transfer to new patient cases—a crucial requirement before any AI tool can be responsibly used in actual clinical practice.

Transparent AI supports clinical decision-making: By pinpointing which specific CpG sites and chromosomal regions influenced each diagnosis, this approach enables pathologists and oncologists to verify that the AI's reasoning aligns with established medical knowledge, building appropriate trust in the system.