Development of a deep learning model for predicting adverse drug reaction using molecular feature of drug and medical data

Abstract

Adverse drug reactions (ADR) refers to the unintended reaction occurring after normal intake of a drug which has harmful effect on a patient’s health and is suspected to have a causal relationship with the drug. ADR has been regarded as one of the major challenges in the modern medical community, not only causing patients to experience serious side effects, but also inducing extra time and expense required for additional treatment. The establishment of big data sources in the clinical field and the advent of deep learning have sparked the development of various health care models which predict patients’ diseases, including ADR with electronic health records (EHR). In recent years, in the healthcare field, studies to overcome the limitations of EHR and improve the performance of predictive model by creating EHR based federate learning model which combines various formats of data have actively been conducted. In this study, we present a new deep learning model using a representative natural language pre-training model, BERT (bidirectional encoder representations from transformers) to comprehensively grasp the medical data of individual patient described in the common data model (CDM) and molecular characteristics of drugs in the prescription record and predict their ultimate impact on the ADR occurrence. Since ADR is caused by a unique combination of individual biological properties, medical conditions, and drugs, the model developed in this study can have a great strength in that it comprehensively considers ADR-related variables using the attention mechanism in the transformer module. Moreover, we introduced a temporal embedding technique to consider the temporal characteristics of EHR and its corresponding correlation with ADR. In this study, we aimed to construct the deep learning model that more accurately predicts ADR, a significant problem in the modern clinical field, by learning CDM-based individual patient medical data (medical description and time records) and molecular features of drugs. In addition, since our model was constructed based on the pre-training model, it will have high utility for multiple hospitals or drug research after combined with a larger amount of CDM data and more diverse drug molecular information.