Towards end-to-end LLM-based censoring-aware survival analysis
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Towards end-to-end LLM-based censoring-aware survival analysis".
Jane: Here's the gist of the paper. Survival analysis is central to medical prediction, yet large language models (LLMs) are rarely used as end-to-end survival models because censoring prevents straightforward supervised fine-tuning Here we present LLMSurvival, a framework that enables censoring-aware survival analysis with unmodified LLMs operating directly on tabular clinical data Pruning redundant reasoning tokens The objective of the work is to address the challenge that LLMs cannot be directly optimized for survival analysis due to the lack of explicit labels caused by censoring The authors present LLMSurvival, a multi-stage framework designed to apply LLMs to survival analysis directly on tabular data Rather than training the LLM to directly predict survival times, the framework optimizes the model for pairwise comparison This approach reformulates survival analysis as a ranking problem, which naturally circumvents the challenges of censoring and allows for the use of all comparable pairs How the training signal is built The framework involves several key stages to build the training signal, starting with data preparation and moving through reformulation and fine-tuning 1.
Tom: Structured covariates are transformed into natural language descriptions using a standardized template, such as “Age is 79; Gender is FEMALE; T-score (NECK AVG) category is Osteopenia” 2.
Jane: Survival prediction is reformulated as a pairwise ranking task where the model learns to predict which subject would experience the event earlier The loss function used is Lrank = −X(i,j)∈P log P(ri > rj), which optimizes relative risk ranking over valid comparable pairs 3.
Tom: During inference, the risk for a test subject is estimated by comparison against a fixed set of anchors from the training cohort using the formula yˆ(s, ak) = (1, if P(rs, rak) > 0.
Title and authors: Tom: The title is "Towards end-to-end LLM-based censoring-aware survival analysis," and the authors are Yishu Wei, Hexin Dong, Yi Lin, Jiahe Qian, Yi Liu, and Yifan Peng.
Jane: They're from Weill Cornell Medicine at Beth Israel Deaconess Medical Center for the data they used in their study on ICU mortality prediction in MIMIC-IV.
Lu: What's interesting is that they are taking unmodified LLMs and operating them directly on tabular clinical data without making any architectural changes to the model itself.
Meng: So, they aren't just using an LLM as a wrapper for a simpler statistical model; they are trying to get the LLM to handle the survival prediction task end-to-end.
Lalam: They are specifically addressing how censoring, which happens when we don't know the exact time of an event because someone drops out or the study ends early, introduces bias when we just discard those subjects.
Tom: So they are building a way to use these powerful language models on complex survival data without having to painstakingly label every single exact time point, which is a big deal for real clinical prediction.
Jane: It makes the whole process accessible because you don't need perfect, fully labeled datasets for survival analysis when you are using this LLMSurvival framework.
The paper's summary: Tom: The main summary of LLMSurvival is that they move away from predicting the actual time to an event and instead focus on pairwise comparison among subjects who are comparable.
Jane: They transform the problem into a ranking task, where the model learns to predict which subject would experience their event sooner than another one.
Lu: The training signal is built through several steps starting with transforming structured clinical covariates into natural language descriptions using a standardized template like "Age is seventy-nine; Gender is FEMALE; T-score (NECK AVG) category is Osteopenia."
Meng: That sounds like they are essentially turning patient charts into descriptive sentences that the LLM can easily process.
Lalam: And then they use a loss function called Lrank, which optimizes the relative risk ranking over valid comparable pairs, meaning it learns how much one subject's risk is higher than another's without needing an absolute time prediction.
Tom: So when they are training, they are optimizing for the relative difference in risk between two patients rather than trying to predict whether patient A dies at three months or patient B at six months.
Jane: That makes sense because that pairwise ranking naturally handles the uncertainty introduced by censoring much better than asking an LLM to guess an absolute time.
The paper's improvements: Tom: They suggest a few key improvements, and one is how they estimate the risk for a new test subject during inference using comparisons against fixed anchors from their training cohort.
Lu: Specifically, the formula they use is that the risk for a test subject is estimated by comparing it against a fixed set of anchor individuals from the training cohort.
Meng: That means when we get an answer for one patient, you don't just get one number; you get how that patient ranks compared to all those reference patients.
Jane: And another improvement is how they handle deployment; they use compact, publicly available base models like Llama three point one-8B-Instruct and Qwen two point five-7B-Instruct, which makes it practical for local deployment without needing expensive external API calls.
Lalam: They also show that the framework is portable because the same general pipeline can be applied to completely different clinical tasks, like acute care versus chronic fracture risk, without needing any major architectural customization.
Tom: So they're not just solving survival analysis; they've shown this framework can be used for diverse clinical prediction problems simply by changing the input data structure slightly.
Conclusion: Jane: To wrap up, the implication here is that we can use unmodified large language models directly on tabular clinical data to perform censoring-aware survival analysis without having to fundamentally alter their architecture.
Tom: The key result is that by reformulating it as a ranking problem, they manage to get a training signal for survival prediction even when the exact event times are missing due to censoring.
Lu: It moves the LLM's role from just being a feature extractor into a comprehensive end-to-end modeling pipeline capable of handling these complex statistical tasks directly on medical records.
Meng: From an engineering standpoint, the fact that it works with models as small as Llama three point one-8B means this could be deployed locally at many different clinics without massive computational overhead for inference.
Lalam: And what's really important is that they provide clinically coherent comparative reasoning outputs, meaning the model can actually generate explanations highlighting meaningful differences between patient profiles based on those risk rankings.
Tom: So, if you want to use LLMs for survival analysis in a practical way right now, the LLMSurvival framework shows you how to do it by focusing on relative risk ranking instead of absolute time prediction.
Jane: That's what they did with "Towards end-to-end LLM-based censoring-aware survival analysis," and it opens up a new path for applying these models in areas where we deal with messy, real-world medical data.
Yishu Wei, Hexin Dong, Yi Lin, Jiahe Qian, Yi Liu, Yifan Peng
Department of Population Health Science, Weill Cornell Medicine
cs.AI
Submitted: 2026-05-25
Updated: 2026-11-10
Code: https://github.com/MIT-LCP/mimic-code
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: Survival analysis is central to medical prediction, yet large language models (LLMs) are rarely used as end-to-end survival models because censoring prevents straightforward supervised fine-tuning…
Key concepts
- Pairwise Ranking Formulation
- Instead of trying to predict an exact survival time (which is hard with censoring), the model learns to rank two subjects against each other. It predicts which subject would experience the event sooner. This ranking approach naturally handles censored data because it only requires comparing comparable pairs, bypassing the need for explicit survival labels.
- Relative Risk Ranking Loss (Lrank)
- The training objective is defined by a loss function that optimizes the relative risk between subjects. It penalizes predictions where the model incorrectly ranks two subjects when they should have a specific relative ordering concerning their event times, effectively training the LLM on comparative risk assessment.
- Anchor Comparison Inference
- During prediction, the model estimates a subject's survival risk by comparing them against a fixed set of 'anchor' subjects from the training data. The risk is determined by calculating how many anchors are ranked as higher risk than the test subject, providing a robust way to estimate survival probability.
- Censoring-Aware Survival Analysis
- This refers to the specific challenge of analyzing survival data where some patients do not experience the event (censored). The framework overcomes this by shifting the focus from absolute time prediction to relative comparison, making it possible for LLMs to perform meaningful analysis on clinical data with missing outcome information.
Terminology
Summary
Survival analysis is central to medical prediction, yet large language models (LLMs) are rarely used as end-to-end survival models because censoring prevents straightforward supervised fine-tuning Here we present LLMSurvival, a framework that enables censoring-aware survival analysis with unmodified LLMs operating directly on tabular clinical data
Pruning redundant reasoning tokens
The objective of the work is to address the challenge that LLMs cannot be directly optimized for survival analysis due to the lack of explicit labels caused by censoring The authors present LLMSurvival, a multi-stage framework designed to apply LLMs to survival analysis directly on tabular data Rather than training the LLM to directly predict survival times, the framework optimizes the model for pairwise comparison This approach reformulates survival analysis as a ranking problem, which naturally circumvents the challenges of censoring and allows for the use of all comparable pairs
How the training signal is built
The framework involves several key stages to build the training signal, starting with data preparation and moving through reformulation and fine-tuning
-
Structured covariates are transformed into natural language descriptions using a standardized template, such as “Age is 79; Gender is FEMALE; T-score (NECK AVG) category is Osteopenia”
-
Survival prediction is reformulated as a pairwise ranking task where the model learns to predict which subject would experience the event earlier The loss function used is Lrank = −X(i,j)∈P log P(ri > rj), which optimizes relative risk ranking over valid comparable pairs
-
During inference, the risk for a test subject is estimated by comparison against a fixed set of anchors from the training cohort using the formula yˆ(s, ak) = (1, if P(rs, rak) > 0. p. 7
Where the paper says it
[INSTANCE 2] Please provide your answer (a or b). 1 Inference via anchor comparison In inference, the risk for a test subject was estimated by comparison against a fixed set of anchors sampled from the training cohort.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Pairwise ranking formulation for censored survival analysis allows training an LLM to
predict which subject would experience the event earlier,
naturally circumventing censoring challenges by optimizing forrelative risk ranking rather than absolute survival time.
-
Inference via anchor comparison estimates risk by calculating
the proportion of anchors for which the test subject was ranked as higher risk,
providing a robust method for estimating survival risk from an LLM output. -
The framework enables local deployment because it uses compact, publicly available base models like Llama 3.1-8B-Instruct and Qwen 2.5-7B-Instruct, mitigating privacy risks by eliminating the need for external API calls.
-
The model is portable because
the same general pipeline was applied to an acute critical care task and a chronic fracture risk task without architectural customization,
demonstrating its ability to handle diverse clinical prediction tasks. -
LLMSurvival can serve as a
comprehensive, end-to-end modeling pipeline
by expanding the LLM's role frommere feature extractors or bottom layers of the overall architecture to a comprehensive, end-to-end modeling pipeline.
-
The framework allows for clinically coherent comparative reasoning by generating explanatory outputs where models highlight
clinically meaningful differences between patient profiles,
suggesting that comparison-based outputs can be accompanied bycoherent, human-readable rationales grounded in known clinical risk factors.
Abstract
Objective: Survival analysis is central to medical prediction, yet large language models (LLMs) are rarely used as end-to-end survival models because censoring prevents straightforward supervised fine-tuning. Here we present LLMSurvival, a framework that enables censoring-aware survival analysis with unmodified LLMs operating directly on tabular clinical data. Materials and Methods: LLMSurvival reformulates time-to-event prediction as pairwise ranking among comparable subjects, and derives test-time risk by aggregating comparisons against anchor individuals from the training cohort. Results: Across two clinical tasks (ICU mortality prediction in MIMIC-IV and fragility fracture prediction in a NewYork-Presbyterian/Weill Cornell Medicine cohort), LLMSurvival improves overall concordance over Cox proportional hazards modeling by 3.1% for ICU mortality and 0.5% for fracture risk, 2.1% on average for ICU mortality and 2.8% for fracture risk over three established deep learning survival models. Discussion: The results show that survival modeling with censoring can be made compatible with LLM fine-tuning through comparison-based reformulation. The framework demonstrates high portability and superior performance over expert curated scores like SAPS-II and FRAX scores across diverse clinical context. Furthermore, the framework supports local deployment, as compact, publicly available base models provide sufficient performance. Conclusion: The LLMSurvival framework serves as a proof of concept for an integrated, censoring-conscious approach to survival analysis via LLMs.
Sources
- Leveraging Large Language Models and Survival Analysis for Early Prediction of Chemotherapy Outcomes
- Learning to rank for censored survival data
Related papers
- Multi-Objective Structured Pruning of LLMs for Latency and Model Size Optimization
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Embeddings for Preferences, Not Semantics
- Can We Trust Item Response Theory for AI Evaluation?
- Diagnosing Task Insensitivity in Language Agents