전자의무기록 데이터 기반 맞춤형 약물 치료를 위한 의료 인공지능 모델 개발 및 데이터베이스 구축

Abstract

Many patients undergo drug therapy for disease treatment. However, due to the unique characteristics of each individual, even when receiving the same drug, the drug response can vary among patients. Consequently, personalized drug therapy that takes into account individual characteristics is necessary, with different drug regimens for each patient. Personalized drug therapy offers the advantage of minimizing drug side effects while maximizing treatment effectiveness. Therefore, both currently used drugs and under-development drugs should be used as personalized medications. In this paper, two studies were conducted to help personalized drug therapy by utilizing electronic medical record (EMR) data from tertiary hospitals. In the first study, we developed and validated a machine learning model for early prediction of the discharge dosage of the anticoagulant drug warfarin. We developed four machine learning models suitable for predicting drug dosage, and through internal validation, we confirmed that the model predictions were more accurate than those of clinical experts. Additionally, we utilized the SHAP (SHapley Additive exPlanations) technique to analyze the key variables that influence the model predictions and explain the model prediction. Finally, we observed significant variability in dosage determination depend on physician’s individual medical experiences, when presented with the same dataset. In contrast, the model's predictive accuracy demonstrated a clinical utility that was twice as high as those of physicians. In the second study, we constructed a novel clinical field-based database by integrating electronic medical records (EMR) and pharmaceutical databases. FDA-approved anticancer agents and associated target gene information was extracted from the Open Targets Platform. We standardized the drug components in both the EMR and Open Targets Platform, and established a linkage between the two databases based on the drugs. As a result, the novel database was included associations between 57 anticancer agents, 60 types of cancer, and 91 genetic mutations. Besides, the database was included additional diagnostic information and genetic test results of patients prescribed with anticancer agents. This integration of data sources allowed for the utilization of both clinical and genetic characteristics of patients in real-world clinical settings, utilizing for personalized cancer treatment. In this study, we have developed two tools that utilize electronic medical record (EMR) data to facilitate personalized drug treatment. First, the machine learning models that predicts the optimal dosage of warfarin can be used as a clinical decision support system to reduce unnecessary treatment duration and contribute to the prevention of drug side effects. Second, the heterogeneous database that integrated both of EMR data and pharmaceutical information databases can be utilized in artificial intelligence-driven drug development.