A Comprehensive Review and Application of Interpretable Deep Learning Model for ADR Prediction

Abstract

Drug safety is a pressing need in today's healthcare. Minimizing drug toxicity and improving the individual’s health and society is the key objective of the healthcare domain. Drugs are clinically tested in laboratories before marketing as medicines. However, the unintended and harmful effects of drugs are called Adverse Drug Reactions (ADRs). The impact of ADRs can range from mild discomfort to more severe health hazards leading to hospitalization and in some cases death. Therefore, the objective of this research paper is to design a framework based on which research papers are collected from both ADR detection and prediction domain. Around 172 research articles are collected from the sites like ResearchGate, PubMed, etc. After applying the elimination criteria the author categorized them into ADR detection and prediction themes. Further, common data sources and algorithms as well as the evaluation metrics were analyzed and their contribution to their respective domains is stated in terms of percentages. A deep learning framework is also designed and implemented based on the research gaps identified in the existing ADR detection and prediction models. The performance of the deep learning model with two hidden layers was found to be optimum for ADR prediction and further, the non-interpretability part of the model is addressed using a global surrogate model. The proposed architecture has successfully addressed multiple limitations of existing models and also highlights the importance of early detection & prediction of adverse drug reactions in the healthcare industry.