LLMSurvival: LLM-Based Censoring-Aware Survival Analysis
A new framework called LLMSurvival enhances large language models' (LLMs) capabilities in analyzing survival data from clinical tables while effectively managing censoring issues. This innovative method emphasizes pairwise comparisons for time-to-event predictions and evaluates risk based on training data references. In two clinical scenarios—one focused on ICU mortality using MIMIC-IV data and the other on fracture risk from NewYork-Presbyterian/Weill Cornell Medicine—LLMSurvival demonstrated a 3.1% improvement in predicting ICU mortality and a 0.5% enhancement for fracture risk compared to traditional Cox proportional hazards models. The framework significantly tackles censoring challenges in survival analysis with LLMs.
Key facts
- LLMSurvival enables end-to-end survival analysis with unmodified LLMs.
- It reformulates time-to-event prediction as pairwise ranking among comparable subjects.
- Test-time risk is derived by aggregating comparisons against anchor individuals from the training cohort.
- Evaluated on ICU mortality prediction in MIMIC-IV and fragility fracture prediction in a NewYork-Presbyterian/Weill Cornell Medicine cohort.
- Improved concordance over Cox proportional hazards by 3.1% for ICU mortality and 0.5% for fracture risk.
- Censoring prevents straightforward supervised fine-tuning of LLMs for survival analysis.
- The framework operates directly on tabular clinical data.
- Published on arXiv with ID 2605.25399.
Entities
Institutions
- NewYork-Presbyterian
- Weill Cornell Medicine
- arXiv