PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Tianqi Xiang, Qixiang Zhang, Xinpeng Ding, Yi Li, Xiaomeng Li

arXiv:2608.16594v1cs.AI

TL;DR

Patient reports remain underexplored for cancer survival prediction, while time-regression LLMs do not directly model concordance-based patient ordering or censored outcomes. CACSurv uses comparative mini-cohort ranking with concordance-aligned supervision and achieves the best C-index across six cancer cohorts, averaging 0.722.

  • Problem

    Patient reports remain underexplored for cancer survival prediction, and independent time regression mismatches concordance-based ordering and right-censored supervision.

  • Method

    CACSurv reformulates report-centric survival prediction as mini-cohort comparative ranking, using concordance-aligned rewards to learn valid orderings under right censoring.

  • Results

    0.722 average C-index; CACSurv achieves the highest C-index on every cohort across six cancer cohorts.

  • Takeaways & Limitations

    CACSurv provides a report-centric survival framework whose predictions align with cohort-level concordance evaluation while using censored outcomes as ranking supervision.

Abstract

from arXiv · show

Cancer survival prediction supports treatment planning, risk stratification, and follow-up management. Existing methods use structured clinical variables, whole-slide images, genomic profiles, or multimodal inputs, while patient reports remain underexplored. We study report-centric survival prediction using reports that organize pathological, clinical, and molecular evidence. Large language models (LLMs) can reason over such reports, but case-wise time regression introduces two mismatches. First, a formulation mismatch arises because survival evaluation depends on ordering comparable patients, whereas independent time predictions do not enforce ranking consistency. Second, a supervision mismatch arises because a censored patient's observed time indicates survival beyond that point and cannot serve as an exact regression target, although it still implies orderings relative to patients who died earlier. To address these mismatches, we propose CACSurv, a Concordance-Aligned Comparative framework for report-centric survival prediction. CACSurv reformulates survival modeling as mini-cohort comparative reasoning, where an LLM predicts relative prognostic orderings. We introduce concordance-aligned rewards derived from comparable relations under right censoring, enabling censored outcomes to provide ranking supervision without exact event-time targets. At inference, Monte Carlo Reference Aggregation compares each patient with sampled references and aggregates positions into a cohort-level ranking. We establish TCGA-SurvReport, a benchmark covering six TCGA cancer cohorts. CACSurv achieves the highest C-index on all six cohorts and an average C-index of 0.722, outperforming the strongest published survival model by 6.5 percentage points and the strongest LLM time-regression baseline by 4.2 percentage points. Our code, models, and dataset will be available at https://github.com/xmed-lab/CACSurv.

Introduction

CACSurv addresses report-centric cancer survival prediction by replacing independent time regression with comparative prognostic ordering aligned to concordance evaluation. It uses right-censoring-aware relational supervision and Monte Carlo Reference Aggregation to produce cohort-level rankings.

  • Motivation: Patient reports organize diagnosis, staging, pathological findings, and molecular characteristics but remain comparatively underexplored for cancer survival prediction.Their information-rich textual interface motivates report-centric survival analysis.
  • Problem: Independent time regression does not explicitly constrain the patient orderings measured by the concordance index.The C-index evaluates whether comparable patients are ordered correctly rather than whether individual predictions are numerically accurate.
  • Problem: Over 60% of the benchmark consists of right-censored cases, whose follow-up times cannot be treated as exact survival targets.Using those times as exact targets introduces incorrect supervision, while excluding censored patients discards prognostic information.
  • Contributions: CACSurv reformulates report-centric LLM survival analysis as mini-cohort prognostic ordering instead of independent absolute-time generation.This aligns the prediction formulation with concordance-based evaluation.
  • Contributions: Concordance-aligned rewards exploit valid partial-order supervision under right censoring, while Monte Carlo Reference Aggregation converts comparative predictions into cohort-level rankings.The approach lets censored patients contribute supervision without treating follow-up times as observed event times or excluding them from training.

Related Work

Prior cancer survival research has extensively studied single-modality approaches based on structured clinical variables and histopathology images, while multimodal methods integrate pathological, genomic, transcriptomic, clinical, and textual evidence. These multimodal methods model cross-modal interactions through co-attention, optimal-transport alignment, and pathway-aware connections.

  • Single-Modality Survival Analysis: Single-modality survival analysis uses statistical or deep models for structured clinical variables and multiple-instance learning to aggregate whole-slide image patch representations.The passage cites Cox (1972), Katzman et al. (2018), Lee et al. (2018), Lu et al. (2021), and Shao et al. (2021) as examples.
  • Multimodal Survival Analysis: Multimodal survival methods combine histopathology with genomic, transcriptomic, clinical, and textual information to capture complementary prognostic evidence.Early approaches model cross-modal interactions using co-attention or optimal-transport-based alignment, while pathway-aware methods connect histological representations with biological pathways.

Methodology

CACSurv reformulates report-centric survival prediction as comparative ranking over mini-cohorts rather than independent time regression. It trains with concordance-aligned rewards under right censoring and infers cohort-level rankings through Monte Carlo Reference Aggregation.

  • Comparative mini-cohort ranking: CACSurv uses mini-cohorts as prediction units and directly learns relative survival orderings among same-cancer-type patients.Training repeatedly exposes the model to inter-patient survival relations instead of predicting absolute survival times or continuous risk scores.
  • Monte Carlo Reference Aggregation: At inference, Monte Carlo Reference Aggregation compares each evaluation patient with randomly sampled same-cancer-type reference patients and averages normalized rank positions.Sorting patients by their aggregated positions produces the final cohort-level survival ordering.
  • Monte Carlo Reference Aggregation: MCRA avoids direct evaluation-patient comparisons and gives every patient a fixed number of comparisons regardless of evaluation cohort size.This supports both small evaluation cohorts and sequential patient arrival.
  • Concordance-aligned rewards: Right-censored patients provide ranking supervision because censoring establishes longer survival than patients who died earlier, without specifying an exact time.CACSurv defines pair correctness rewards for pairs and triplet concordance rewards over comparable relations in larger mini-cohorts.
  • Concordance-aligned rewards: The unified concordance-aligned reward selects pair or multi-patient concordance rewards and supplies the GRPO training signal.For cohorts of at least three patients, the reward aggregates ordering consistency across all within-cohort comparable relations.

Experiments · Dataset Construction and Cohort Statistics

TCGA-SurvReport is a report-centric survival dataset spanning six TCGA cancer cohorts and aligning OCR-derived pathology text, clinical records, and molecular profiles by patient identifier. The construction pipeline standardizes heterogeneous reports while retaining patients with valid unified reports and survival records, with survival labels kept separate from report content.

  • Dataset Construction and Cohort Statistics: TCGA-SurvReport covers six TCGA cancer cohorts: BLCA, BRCA, COADREAD, HNSC, LUAD, and STAD.The benchmark is designed for report-centric survival prediction.
  • Dataset Construction and Cohort Statistics: Each patient record combines OCR-derived pathology text, clinical records, and molecular profiles.These sources provide complementary pathological, demographic, clinical, and molecular information.
  • Dataset Construction and Cohort Statistics: The three data sources are aligned using TCGA patient identifiers.Patient identifiers provide the linkage across pathology, clinical, and molecular information.
  • Dataset Construction and Cohort Statistics: Released OCR text contains heterogeneous formats, duplicated content, administrative statements, and detailed specimen-processing descriptions.These characteristics motivate report cleaning and standardization.
  • Dataset Construction and Cohort Statistics: Qwen2.5-72B-Instruct uses a fixed extractive prompt to remove irrelevant content and organize retained information into standardized patient reports.The model is prohibited from adding diagnoses, medical facts, prognostic interpretations, or clinical implications absent from source records.
  • Dataset Construction and Cohort Statistics: Observed time and vital status are extracted separately as survival labels and excluded from unified patient reports.This separates survival supervision from the report text used as model input.
  • Dataset Construction and Cohort Statistics: Patients are retained only when both a valid unified report and an available survival record are present.Figure 3 summarizes cohort statistics across the six cancer types.

Baselines

The study compares CACSurv against representative WSI-based, multimodal, report-centric, and LLM time-regression survival baselines. These baselines span WSI-only, histology–molecular, histology–report–molecular, tabular, biomedical-language-model, and instruction-tuned LLM settings.

  • WSI-based and multimodal baselines: The comparison includes WSI-based and multimodal survival models spanning WSI-only, histology–molecular, and histology–report–molecular settings.Representative models include TransMIL, TITAN, MCAT, MOTCat, SurvPath, PS3, PAMoE, CIMA, SlotSPE, and DPSurv.
  • Report-centric baselines: Report-centric baselines comprise Tabular-CoxPH, Tabular-DeepHit, PubMedBERT-DeepSurv, and BioMistral-CoxPH.These methods represent tabular and biomedical-language-model approaches for report-centric survival comparison.
  • LLM time-regression baselines: LLM time-regression baselines use zero-shot Qwen2.5-7B/72B-Instruct and supervised fine-tuning with Qwen2.5-7B-Instruct.The zero-shot and supervised settings provide LLM-based time-regression comparisons.

Implementation and Evaluation Protocol

CACSurv and all baselines are compared under identical patient-level five-fold cross-validation, with performance summarized by mean±standard deviation C-index. Training uses comparable pair and triplet mini-cohorts with reports as the only model inputs, while references are drawn from training patients at inference.

  • Evaluation protocol: All methods use identical patient-level five-fold cross-validation splits and cohorts, with models trained only on each fold’s training patients.Baseline results are rerun under this unified evaluation protocol.
  • Evaluation protocol: Performance is reported as mean±standard deviation C-index across the five cross-validation folds.Table 1 reports the main TCGA-SurvReport results under this 5-fold protocol.
  • Input processing: WSI methods use CONCHv1.5 patch features, PS3 uses PLIP, and tabular or report-embedding baselines follow their respective input-processing implementations.Tabular-CoxPH and Tabular-DeepHit use structured report variables, whereas PubMedBERT-DeepSurv and BioMistral-CoxPH use model-generated report embeddings.
  • Training protocol: Each fold samples 1,000 comparable pairs and 1,000 triplets per cancer type and optimizes the model for one epoch.Every triplet contains at least one comparable patient pair.
  • Training protocol: Observed time and vital status constrain cohort sampling and compute concordance-aligned rewards, while only patient reports are provided to the model.This prevents outcome variables from being used as direct model inputs.
  • Inference protocol: At inference, the training patients form the reference bank for comparing test patients.The supplied passage specifies the reference-bank construction but truncates the remainder of the inference procedure.

Comparisons with State-of-the-Art Methods

CACSurv delivers the strongest survival-ranking performance across all six cancer cohorts, surpassing published survival models and direct LLM time-regression baselines. Its comparative formulation outperforms the strongest alternatives by substantial margins.

  • Overall comparison: CACSurv achieves the highest C-index on every cohort and the best average C-index of 0.722.Its simplified pair-only variant ranks second, with an average C-index of 0.702.
  • Published survival models: 6.5 percentage points: CACSurv exceeds SlotSPE, the strongest published survival model, whose average performance is 0.657.CACSurv also consistently outperforms evaluated WSI-based, multimodal, and conventional report-centric survival models.
  • LLM time-regression baselines: 4.2 percentage points: CACSurv with a 7B backbone outperforms Qwen2.5-72B Time-ZS, the strongest LLM-based absolute-time regression baseline.This result indicates that comparative prediction is more effective for this task than direct absolute-time regression.

Ablation on Pair and Mixed Mini-Cohorts

Mixed pair-and-triplet mini-cohorts improve survival-ranking performance over pair-only training, with the best results when mixed mini-cohorts are used during both training and inference.

  • Mixed Mini-Cohort Ablation: 0.722 average C-index is achieved with mixed pair-and-triplet mini-cohorts in both training and inference, compared with 0.702 for pair-only training under pair inference.Mixed mini-cohorts use 1,000 pairs and 1,000 triplets, while pair-only training uses 2,000 pairs.
  • Mixed Mini-Cohort Ablation: 0.019 average C-index improvement results from pair-and-triplet training under pair-and-triplet inference, increasing performance from 0.703 to 0.722.The total number of training mini-cohorts, backbone, data splits, and training epochs remain unchanged.
  • Mixed Mini-Cohort Ablation: 0.013 average C-index improvement results from pair-and-triplet training under pair inference, increasing performance from 0.702 to 0.715.The results indicate that triplet relations provide complementary comparative supervision.

Ablation on Reward-based Optimization

Under identical settings, CACSurv’s concordance-aligned reward optimization with GRPO consistently outperforms Rank-SFT across all six cohorts. Rank-SFT instead directly fine-tunes the model to generate target rankings.

  • Optimization comparison: CACSurv consistently outperforms Rank-SFT across all six cohorts when only the optimization method varies.Table 3 reports three representative cohorts and the six-cohort average for compactness.
  • Optimization comparison: Rank-SFT directly fine-tunes the model to generate the target ranking, whereas CACSurv optimizes concordance-aligned rewards with GRPO.

Sensitivity to Number of Reference Comparisons · CACSurv Rollout Example

CACSurv’s performance improves with more reference comparisons and largely converges at K = 40, which balances accuracy and complexity. Its rollout converts report evidence into explicit pairwise survival rankings through prognostic-factor identification and patient comparison.

  • Sensitivity to Number of Reference Comparisons: 0.723 C-index is achieved at K = 80 under CACSurv’s default Train.Mix and Eval.Mix setting.Increasing K from 40 to 80 improves the C-index from 0.722 to 0.723.
  • Sensitivity to Number of Reference Comparisons: Performance improves consistently as the number of reference comparisons K increases.The trained models and all other inference settings remain unchanged for fairness.
  • Sensitivity to Number of Reference Comparisons: Performance largely converges at K = 40.This convergence motivates using K = 40 for inference.
  • Sensitivity to Number of Reference Comparisons: K = 40 is adopted as an acceptable accuracy-complexity tradeoff.The default setting uses Train.Mix and Eval.Mix.
  • CACSurv Rollout Example: The rollout example uses a pair mini-cohort to demonstrate CACSurv’s comparative prediction process.Figure 4 presents the example for two patients.
  • CACSurv Rollout Example: CACSurv first identifies patient-specific prognostic factors from the reports.These factors provide the evidence used in the subsequent comparison.
  • CACSurv Rollout Example: The model compares the two patients based on the extracted evidence.The comparison precedes the final survival ranking.
  • CACSurv Rollout Example: The rollout finally produces a survival ranking from the comparative evidence.The example shows how CACSurv converts report evidence into an explicit comparative prediction.

Conclusion

CACSurv reformulates report-centric cancer survival prediction as comparative prognostic ordering within mini-cohorts, using concordance-aligned rewards that exploit valid relations under right censoring. Monte Carlo Reference Aggregation produces cohort-level rankings, and CACSurv achieves the best performance across six TCGA cohorts with an average C-index of 0.722.

  • Framework: CACSurv predicts prognostic orderings within mini-cohorts and optimizes concordance-aligned rewards from valid comparable relations under right censoring.This addresses formulation and supervision mismatches in conventional LLM-based time regression.
  • Framework: CACSurv rollouts identify patient-specific prognostic factors, compare discriminative evidence across patients, and derive a ranking decision.The predicted survival ordering is returned in the answer section.
  • Inference: Monte Carlo Reference Aggregation converts local comparative predictions into stable cohort-level survival rankings.The aggregation operates after the model produces comparative predictions within sampled reference mini-cohorts.
  • Evaluation: 0.722 average C-index: CACSurv achieves the best performance among evaluated methods on every cohort across six TCGA cancer cohorts.The benchmark, TCGA-SurvReport, uses identical patient-level five-fold cross-validation splits.
Loading 2608.16594v1…