Survprompt
A prompting framework that converts structured patient covariates into free-text clinical vignettes for input to a large language model, enabling zero-shot survival time prediction. It is used to evaluate the capability of large language models on medical prognosis tasks.
Timeline (last 90 days)
The Survprompt framework converts structured patient covariates into free-text clinical vignettes, which are then fed to a pre-trained LLM to predict survival in a zero-shot manner.
Feature ablation showed that the LLM's prioritization of clinical variables is highly similar to that of specialized survival models.
LLMs showed unstable performance across different cancer types and institutions, and had poor ability to distinguish between high- and low-risk patients (low c-index).
All Facts (3)
The Survprompt framework converts structured patient covariates into free-text clinical vignettes, which are then fed to a pre-trained LLM to predict survival in a zero-shot manner.
50%UnverifiedFeature ablation showed that the LLM's prioritization of clinical variables is highly similar to that of specialized survival models.
50%UnverifiedLLMs showed unstable performance across different cancer types and institutions, and had poor ability to distinguish between high- and low-risk patients (low c-index).
50%