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

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Survival analysis and patient risk evaluation are critical in healthcare.
• Deep learning models offer personalized prognosis but require extensive data, often limited in clinical settings.
• Challenges include manual data entry, missing values, and non-numerical covariates, hindering model performance.

Purpose of the Study:

• To propose a novel approach using large language models (LLMs) to generate synthetic patient data for survival analysis.
• To address challenges of limited data, missing values, and non-numerical covariates in healthcare datasets.
• To develop a prognostic model capable of risk stratification using human-like sentences.

Main Methods:

• Utilized large language models (LLMs) to create comprehensive sentences from patient information, including missing data.
• Simulated natural variability in clinical documentation through random masking.
• Developed a simple network trained on LLM-generated synthetic data for survival analysis.

Main Results:

• The proposed method effectively generated synthetic data that enabled a trained network to achieve comparable results to previous survival analysis research.
• Performance metrics on FLCHAIN, METABRIC, and SUPPORT datasets demonstrated the model's efficacy (C-index: 0.857, 0.690, 0.985; IBS: 0.108, 0.188, 0.213).
• The approach successfully stratified risk groups using intuitive, human-like sentences.

Conclusions:

• LLM-generated synthetic data offers a viable solution to data scarcity and heterogeneity in healthcare for survival analysis.
• This method simplifies data preprocessing by eliminating the need for conversion of non-numerical data and handling missing values.
• The developed prognostic model provides an intuitive and effective tool for patient risk stratification.