Survprompt: Leveraging Large Language Models for Survival Analysis
Summary:Survprompt is an innovative framework that converts structured patient covariates into free-text clinical vignettes and leverages pre-trained large language models (LLMs) for zero-shot survival prediction. Experiments on multi-institutional cancer datasets show that frontier LLMs achieve competitive performance in predicting individual survival times for certain cancer types, though their accuracy varies across cancer types and institutions. This research highlights the potential of LLMs in medi
Background and Motivation
Survival analysis is a critical tool in medical risk assessment, predicting the time-to-event outcomes for patients. Traditional methods like Random Survival Forests (RSF) require substantial data and specific training, while Large Language Models (LLMs) offer new possibilities for survival analysis due to their strong text understanding and generation capabilities.
The Survprompt Framework
Survprompt operates through the following steps:
- Data Conversion: Transforming structured patient covariates into free-text clinical vignettes.
- Zero-shot Prediction: Utilizing pre-trained LLMs to predict survival times based on the converted text.
Experimental Results
- Performance: Experiments on the MSK-CHORD and Providence St. Joseph Health Network datasets show that Survprompt achieves competitive performance in predicting individual survival times for certain cancer types, with GPT-5.6-Sol achieving a cMAE lower than RSF for prostate cancer.
- Limitations: The accuracy of Survprompt varies significantly across cancer types and institutions, and it struggles to discriminate between high- and low-risk patients (lower c-index).
Technical Highlights
- Zero-shot Learning: The ability to perform survival prediction without specific training data demonstrates the generalization capabilities of LLMs.
- Clinical Variable Prioritization: LLMs prioritize clinical variables related to survival, similar to traditional survival models.
- Multi-institutional Datasets: The use of multi-institutional datasets enhances the reliability and generalizability of the research findings.
Industry Impact and Developer Recommendations
- Medical Applications: Survprompt showcases the potential of LLMs in medical prognosis, providing new tools for clinicians and researchers.
- Cautions: In clinical settings, predictions from LLMs should be treated with caution, especially regarding variability across cancer types and institutions.
- Future Directions: Further research should focus on improving the consistency of LLM predictions across different cancer types and institutions and exploring their application in multimodal data processing.
Conclusion
Survprompt demonstrates the potential of LLMs in survival analysis but also highlights their limitations. Future studies should aim to enhance the consistency of LLM predictions and explore their application in multimodal data processing.
— END —Source: ArXiv NLP/LLM (cs.CL) (2026-10-01)
Tags: #Large Language Models #Survival Analysis #Medical Prognosis #Zero-shot Learning #Clinical Applications
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments