CACSurv: Concordance-Aligned Comparative Learning with Large Language Models for Cancer Survival Prediction

Details

Standard cancer survival models predict absolute event times for individual patients, which creates two problems: survival evaluation depends on relative ordering of patients (concordance), and censored observations cannot provide exact regression targets. CACSurv addresses both by reframing the task as comparative reasoning — an LLM ranks relative prognosis within mini-cohorts rather than predicting absolute survival times. The framework introduces concordance-aligned rewards derived from pairwise patient comparisons under right censoring, enabling censored outcomes to contribute ranking supervision without requiring exact event times. At inference, Monte Carlo Reference Aggregation compares each patient against sampled references, then aggregates positions into a cohort-level ranking. Evaluated on TCGA-SurvReport, a new benchmark the authors introduce covering six TCGA cancer cohorts, CACSurv achieves a C-index of 0.722 — outperforming the strongest published survival model by 6.5 percentage points and LLM time-regression baselines by 4.2 points. Code, models, and the benchmark dataset will be released. The practical implication is that patient reports (pathological + clinical + molecular) are a viable, underexplored input modality for survival prediction.