SAGEAgent reduces multimodal survival prediction burden by 55%

Researchers propose SAGEAgent, a self-evolving LLM-based clinical agent that actively decides which diagnostic modalities to acquire for cancer patients, balancing predictive accuracy against clinical invasiveness. The system formulates modality selection as a sequential decision problem, utilizing clinical tools, episodic memory, and semantic memory to reason about each patient's evolving diagnostic state.

• SAGEAgent reasons through clinical tools that translate numerical predictions into text.
• It employs episodic memory to retrieve similar past cases and semantic memory for reusable decision patterns.
• Experiments on a glioma cohort combining TCGA-LGG, TCGA-GBM, and BraTS with four diagnostic modalities show competitive survival prediction accuracy.
• The approach reduces the average acquisition burden by 55% compared to methods assuming all modalities are available.

This work addresses the limitation of current multimodal survival methods that either assume complete data availability or passively handle missing data without justifying the acquisition of subsequent modalities.