Quantifying Cross-Modal Interactions in Multimodal Glioma Survival Prediction via InterSHAP: Evidence for Additive Signal Integration

Multimodal AI doesn't always work through synergy: The study challenges a common assumption that combining different data types (like medical images and genetic information) creates powerful interactions. Instead, the research found that better-performing models simply add complementary information together rather than discovering hidden synergies between data types.

Better performance comes from complementary signals, not complexity: Across all four fusion architectures tested, the breakdown remained consistent—whole-slide images contributed ~40%, RNA data ~55%, and cross-modal interaction only ~4%. This suggests that improving multimodal AI for cancer prognosis may not require increasingly complex architectures, but rather better ways to combine different information sources.

A new tool for auditing AI medical models: The adapted InterSHAP method provides a practical way to measure and compare how different AI architectures actually use multimodal data in survival prediction. This transparency tool could help clinicians and developers understand whether a model's improvements come from genuine synergy or simply from better data aggregation.

Implications for privacy-preserving AI deployment: Since the findings show that additive signal integration outperforms complex fusion, this supports federated learning approaches where different data types (images, genetic data) could be processed separately and combined later—reducing privacy risks in sensitive medical AI applications.

Study validates on real cancer data: The research used 575 glioma patient cases from The Cancer Genome Atlas, providing evidence-based findings rather than theoretical predictions about how multimodal medical AI actually performs in practice.