Advancing Survival Analysis with Large Language Models: A Solution to Data Scarcity and Missing Information in...

Survival analysis and patient risk evaluation are vital tasks in healthcare. Deep learning models promise personalized patient prognosis, but they require large amounts of training data for good results. In healthcare, this presents a particular challenge as most datasets, especially tabular data, are manually entered by clinicians during trials. Missing data poses another significant hurdle. While deleting incomplete rows might seem logical, this further reduces the already limited sample size. A popular alternative is data imputation-estimating missing values using the dataset's distribution-but this requires several assumptions to be valid and useful. Healthcare data also frequently contains non-numerical covariates like text descriptors, which require careful conversion to avoid introducing bias and noise.We propose a novel approach that sidesteps these issues by using large language models (LLMs) to create comprehensive sentences containing all relevant subject information. This approach offers several key benefits: patient data can be added directly without conversion, missing data becomes a non-issue as descriptions can be as detailed as needed, and the natural variability of clinical documentation is simulated through random masking. We demonstrate this method's effectiveness by creating a simple network trained on our synthetic data that achieves comparable results to previous survival analysis research on the FLCHAIN, METABRIC, and SUPPORT datasets. Importantly, our approach is inherently intuitive, presenting a prognostic model that can stratify risk groups using human-like sentences. C-index: 0.857, 0.690, 0.985, IBS: 0.108, 0.188, 0.213.