Presentation Information
[1Yin-B-04]Predicting Treatment Outcomes for Breast Cancer Using Large-Scale Electronic Health Records and Large Language Models
〇Chisato Kamiya1, Ami Sakane1, Kenshi Kawaguchi1, Yuuki Hashimoto2, Hiromasa Horiguchi2, Yoshinobu Kano1 (1. Shizuoka University, 2. National Hospital Organization)
Keywords:
Medical AI,Cancer Treatment
We constructed a model to predict the therapeutic effects of pharmacotherapy for breast cancer using large-scale electronic health record (EHR) data.
We employed two large language models: the general-purpose Llama 3.1 and Japanese medical language specialized SIP-jmed-llm-3.
From EHR texts, we automatically extracted three response categories: “response,” “stable disease,” and “progressive disease.”
The task was designed to predict the next radiology assessment result based on a reference date on which a radiological examination was performed. Features included medication administration data between the reference date and the subsequent examination, as well as laboratory test results up to the reference date, prior medication records, patient demographic information, and breast cancer histopathological data. These features were used to train a classification model with XGBoost. Analysis of feature importance confirmed that the radiology assessment result on the reference date, breast cancer histopathological information, and various anticancer drugs contributed significantly to the prediction.
We employed two large language models: the general-purpose Llama 3.1 and Japanese medical language specialized SIP-jmed-llm-3.
From EHR texts, we automatically extracted three response categories: “response,” “stable disease,” and “progressive disease.”
The task was designed to predict the next radiology assessment result based on a reference date on which a radiological examination was performed. Features included medication administration data between the reference date and the subsequent examination, as well as laboratory test results up to the reference date, prior medication records, patient demographic information, and breast cancer histopathological data. These features were used to train a classification model with XGBoost. Analysis of feature importance confirmed that the radiology assessment result on the reference date, breast cancer histopathological information, and various anticancer drugs contributed significantly to the prediction.