From Hazard Functions to Language Space: Cox-Supervised Distillation of Survival Risk into a Large Language Model

2026-06-08 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

A recent investigation explores transferring time-to-event risk information, estimated by a Cox proportional hazards model, into a generative large language model. Researchers propose a text-based survival modeling pipeline where structured clinical covariates are converted into text prompts. A Qwen-based large language model is then fine-tuned to generate patient-specific survival risk, using Cox model predictions as the training target. Across datasets including GBSG2, ACTG320, and WHAS500, the fine-tuned model achieves competitive held-out discrimination and calibration. This performance is notable given its training as a text-generation task, rather than with a conventional survival-analysis loss. Further analysis using t-SNE visualizations of the model's hidden states reveals smooth risk gradients in latent space, indicating the model represents survival risk as a continuous structure. These findings suggest large language models can internalize survival-risk structure and support calibrated prediction, offering a path towards time-to-event reasoning in language models.

Key takeaway

For Machine Learning Engineers developing predictive models in healthcare, this research suggests a novel approach to integrate survival analysis into large language models. You can fine-tune generative LLMs like Qwen with Cox model outputs, converting structured clinical data into text prompts. This method allows LLMs to internalize complex time-to-event risk structures, potentially simplifying deployment and enabling more intuitive, language-based risk reasoning in clinical applications. Consider exploring this distillation technique to enhance your LLM's predictive capabilities for survival outcomes.

Key insights

Large language models can internalize and predict survival risk from structured clinical data via text-based distillation.

Principles

• Cox model predictions can serve as LLM training targets.
• LLMs can represent continuous survival risk in latent space.
• Text generation tasks can yield calibrated survival predictions.

Method

Convert structured clinical covariates into text prompts, then fine-tune a Qwen-based LLM using Cox model predictions as training targets for text-based survival risk generation.

In practice

• Use text prompts for structured clinical data input to LLMs.
• Distill traditional survival model outputs into LLMs.
• Analyze LLM latent space for risk gradient insights.

Topics

• Large Language Models
• Survival Analysis
• Cox Proportional Hazards
• Clinical Covariates
• Time-to-Event Prediction
• Model Distillation
• Qwen

Best for: NLP Engineer, AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• From Hazard Functions to Language Space: Cox-Supervised Distillation of Survival Risk into a Large Language Model — Large language models can internalize and predict survival risk from Cox models via text-based distillation.
• MILM: Large Language Models for Multimodal Irregular Time Series with Informative Sampling — LLMs can effectively process multimodal irregular time series by encoding sampling patterns and values in XML.
• Linguistic Symptoms: Augmented Generation by Retrieval and Reasoning in LLMs under Portuguese-English Variation in Medical Contexts — Linguistic variation significantly impacts smaller LLMs in medical contexts, with larger models showing greater cross-language robustness.
• ClinicalMC: A Benchmark for Multi-Course Clinical Decision-Making with Large Language Models — Evaluating LLMs in evolving multi-course clinical scenarios is crucial for real-world healthcare deployment.
• A survey on large language models in biology and chemistry — LLMs are transforming molecular sciences by treating biological and chemical systems as structured languages.
• Predicting Causal Effects from Natural Language Queries using Structured Representations — Finetuning and structured representations significantly enhance LLM performance in predicting causal effects from natural language.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.