LLMs show promise for breast cancer prognosis

Fine-tuned LLMs were evaluated as flexible tools for breast cancer survival analysis, a clinical prognosis method used to estimate patient outcomes over time. The work targets a persistent challenge in healthcare modeling: traditional statistical approaches can struggle with complex, high-dimensional data, while access to real clinical records is limited by privacy constraints.

The framework used a synthetic dataset designed to emulate a national population-based cancer registry and comprising 60,000 breast cancer patients after feature engineering and data imputation. Researchers compared Cox regression and gradient boosting with fine-tuned LLMs spanning encoder-only, decoder-only, and encoder-decoder architectures, using survival metrics that account for censoring.

The strongest models were then tested on a real-world cohort of 183,304 patients from Dutch cancer registries. Models trained on synthetic data and real data both maintained strong performance in retrospective evaluation settings, but results declined under standard inference conditions when survival status was unavailable. The findings support synthetic registry data as a privacy-preserving route for model development in data-constrained clinical settings.