Predicting Early-Onset Colorectal Cancer with Large Language Models

Area of Science:

• Oncology
• Artificial Intelligence
• Medical Informatics

Background:

• Early-onset colorectal cancer (EoCRC) incidence is increasing in individuals under 45.
• This demographic is below current national cancer screening age guidelines.
• Predictive models are needed to identify at-risk individuals for EoCRC.

Purpose of the Study:

• To evaluate machine learning (ML) and large language models (LLMs) for predicting early-onset colorectal cancer (EoCRC).
• To compare the predictive performance of various ML models against advanced LLMs.
• To utilize patient journey data within six months preceding diagnosis for EoCRC prediction.

Main Methods:

• Retrospective analysis of 1,953 colorectal cancer (CRC) patients from US health systems.
• Application and comparison of 10 distinct machine learning models.
• Utilized a fine-tuned large language model (LLM) incorporating patient conditions, lab results, and observations.

Main Results:

• The fine-tuned LLM demonstrated superior performance in predicting EoCRC.
• Achieved an average sensitivity of 73% for EoCRC prediction.
• Achieved an average specificity of 91% for EoCRC prediction.

Conclusions:

• Fine-tuned LLMs show significant potential for early detection of colorectal cancer in younger populations.
• LLMs can effectively analyze diverse patient data for cancer risk prediction.
• This approach may aid in developing targeted screening strategies for early-onset colorectal cancer.