CACSurv: LLM-Based Cancer Survival Prediction from Patient Reports

A new research paper introduces CACSurv, a method for cancer survival prediction that leverages large language models (LLMs) to analyze patient reports. The study, available on arXiv (2608.16594), addresses two key mismatches in existing approaches: a formulation mismatch, where survival evaluation relies on ordering comparable patients but independent time predictions lack ranking consistency, and a supervision mismatch, where censored patients' observed times only indicate survival beyond that point, not exact regression targets. CACSurv employs concordance-aligned comparative learning to align LLM predictions with survival ordering. The paper highlights the underexplored use of patient reports, which organize pathological, clinical, and molecular evidence, for survival prediction. The proposed method aims to improve treatment planning, risk stratification, and follow-up management by providing more accurate survival estimates. The research is part of ongoing efforts to integrate LLMs into medical AI, particularly for oncology. The paper was announced on arXiv with the ID 2608.16594.