Beyond classical models: LLM-driven survival analysis for breast cancer prognosis using European cancer registry data

Survival analysis is a fundamental tool in clinical prognosis, yet traditional statistical models often struggle to capture complex, high-dimensional relationships in modern healthcare data. Recent advances in Large Language Models (LLMs) offer new opportunities for flexible and context-aware modeling. At the same time, access to real-world clinical data remains restricted due to privacy constraints, motivating the use of synthetic datasets as a privacy-preserving alternative for model development and evaluation.

We propose a survival analysis framework based on fine-tuned LLMs, evaluated on a large-scale synthetic breast cancer dataset derived from European cancer registry data. A synthetic dataset emulating a national population-based cancer registry and comprising 60,000 breast cancer patients was used following feature engineering and data imputation. We compared traditional survival analysis methods, including Cox regression and gradient boosting, with a range of fine-tuned LLMs representing encoder-only, decoder-only, and encoder-decoder architectures. Model performance was evaluated using standard survival analysis metrics accounting for censoring. To assess generalizability, the best-performing models were deployed on a real-world cohort of 183,304 patients from Dutch cancer registries.

The proposed LLM-based models demonstrate competitive performance across multiple survival metrics, with consistent differences observed across model architectures. When applied to real-world registry data, models trained on synthetic data and those trained on real data retained strong performance in retrospective evaluation settings, while performance decreased under standard inference conditions where survival status was unavailable. This study provides large-scale empirical evidence that models trained on synthetic cancer registry data can generalize to real-world populations, achieving strong performance in retrospective evaluation settings and moderate performance under standard inference conditions.

This study demonstrates that LLMs can serve as flexible models for survival analysis. The use of high-fidelity synthetic data enables privacy-preserving model development while maintaining strong predictive performance. These findings support the integration of Generative AI methods into survival modeling pipelines, particularly in data-constrained clinical settings.