Predicting the signals of adverse drug reactions on the basis of medical data

Abstract

Multiple drugs are often prescribed for the better treatment of a disease. While the use of multiple drugs is to improve the overall therapeutic efficacy, it can also lead to unintended adverse drug reactions (ADRs) from drug-drug interactions (DDIs). More than 30% of ADRs have been reported to be caused by DDIs. Because of the importance of DDIs and ADRs, many computational models have been developed to predict DDIs and/or ADRs. However, these computational models mostly make predictions on the basis of molecular features of drugs without considering clinical factors, which consequently does not allow predictions in a person-specific manner. To address this problem, we present a machine learning model that predicts the signals of ADRs potentially caused by DDIs on the basis of medical data. Medical data used for model development come from MIMIC-IV, and include age, ethnicity, gender, patients’ disease information (i.e., ICD-10) and information on drugs administered (e.g., molecular structure and dose). The medical data were subjected to extensive preprocessing based on a domain knowledge, and multiple machine learning methods were examined to predict the ADR signals using the preprocessed dataset. ADR signals are captured by classifying whether each of the selected laboratory measurements (e.g., hematocrit, creatinine, and hemoglobin) is normal. The best-performing model showed the average AUROC of more than 0.8. The machine learning model developed in this study can be useful for examining potential harms when considering multiple drugs, including anticancer drugs.