Prediction of Treatment Outcomes and Imaging Results for Colorectal Cancer Using Large-Scale Electronic Medical Records

Abstract

In this study, we aimed to predict the treatment efficacy in cancer therapy using examination and medication information, as well as basic patient information extracted from electronic medical records (EMRs) of colorectal cancer patients. As an indicator of treatment efficacy, we focused on imaging examination results, which were extracted from the text in the EMRs using a large language model. We constructed an EMR dataset of 7,257 colorectal cancer patients collected from 22 institutions and conducted experiments using XGBoost. The results showed that we achieved an accuracy of 61.20% for the three-class classification task (response/stable/progression) and 79.92% for the two-class classification task (response or stable/progression). Analysis of important features suggested the influence of liver function-related characteristics.