Jeon, J.; Hong, E.; Kim, J.-Y.; Lee, S.; Kim, H. U.

Various computational models have been developed to understand the physiological effects of drug-drug interactions, which can contribute to more effective drug treatments. However, they mostly focus on interactions of only two drugs, and do not consider the patient information. To address this challenge, we use publicly available electronic health record (EHR), MIMIC-IV, to develop machine learning models that predict the physiological effects of two or more drugs. This study involves extensive preprocessing of laboratory measurement data, prescription data and patient data. The resulting machine learning models predict potential abnormalities across 20 selected measurement items (e.g., concentrations of metabolites and blood cells) in the form of a sentence. The models showed an average AUROC of 0.75, and age, specific active pharmaceutical ingredients, and gender appeared to be the most influential features. The model development process showcased in this study can be extended to other measurement items for a target EHR.