From Predictions to Conversations: An Agentic Framework for Interrogatable and Interpretable Survival Analysis
Abstract
Purpose
Survival models based on whole-slide images (WSIs) often function as "black boxes," hindering clinical adoption despite high accuracy. Clinicians require granular explanations to identify which tissue components drive prognostic risk. We present an agent-based survival analysis framework that enables tissue-specific interpretability and interactive, natural-language exploration of prognostic signals within complex pathology landscapes.
Methods
WSIs were partitioned into tiles and assigned tissue-type labels using a pre-trained vision-language foundation model (CONCH). Tiles were embedded with foundation model UNI2 and organized into tissue-specific graphs to capture micro-architectural relationships and global spatial context. A graph-based survival model was trained using a Cox proportional hazards objective. An agentic interface was integrated, allowing users to pose natural-language queries to trigger targeted analyses and generate visual representations of the tissue regions. This allows for isolating specific components, excluding regions of interest, or comparing risk contributions across different histological compartments.
Results
The framework achieved a peak concordance index (C-index) of 0.601 ± 0.009 (Necrosis + Adipose Tissue) on the TCGA-BLCA cohort, performing competitively with slide-level baselines while providing superior granularity. Analysis revealed significant heterogeneity in prognostic signals; notably, perivesical adipose tissue alone demonstrated a higher C-index (0.591 ± 0.007) than invasive urothelial carcinoma alone (0.580 ± 0.006). Agent-driven queries successfully localized risk signals, providing consistent updates to predictions when subsets were isolated. The agent synthesized these findings into clinically meaningful summaries, translating abstract graph feature importance into reasoning aligned with established staging criteria (e.g., extravesical extension).
Conclusion
This framework transforms static WSI survival predictions into an interactive decision-support tool. By revealing that extra-tumoral regions like perivesical fat can carry more prognostic weight than the tumor itself, the system aligns AI outputs with pathologist intuition. By pinpointing high-risk histological compartments, this framework provides multidisciplinary tumor boards with objective evidence to tailor local-regional treatment strategies, such as radiotherapy escalation.