Adversarial LSTM-Based Sequence-to-Sequence Model for Drug-Related Reactions Understanding

Abstract

Social networks become an extensive source used to automatically report treatment and drug issues. Few approaches have been proposed in this matter, namely multi-word medical concept identification and drug-related descriptions that denoted vast connections for understanding patients’ shared experiences. Therefore, the paper explores the problem of mutual forms of drug interactions including beneficial drug effects. We developed an LSTM-based encoder–decoder model that leverages and memorizes concepts’ details and relationships to form a drug description that compactly with highly informed embeddings seeks multi-word drug-related descriptions from online generated narratives. Meanwhile, we apply adversarial training to regularize sequential embeddings of distributed biomedical n-grams representation from controlled medical vocabulary such as PubMed. It consists of a semi-supervised featuring medical concepts from online generated narratives at sentence-level, especially to create a feature descriptor to select possible drug-related descriptions. The model achieves an accuracy of 86% in assessing false positives and negatives of patients’ statements. Whereby the model shows a well-understanding of real-world medication-related descriptions with significant improvements over baseline neural network approaches.