The Discovery Of Potential Adverse Drug Reactions Based On Text Mining

With the development of biomedical technologies, a growing number of drugs flow into the market. On the one hand, drugs can treat human diseases, improve human health level and even extend human life. On the other hand, the adverse drug reactions make serious damages to human health in some conditions, and some of them even are fatal. Excepting causing damages to human bodies, adverse drug reactions can cause enormous economic loss. How to find adverse reactions for drugs timely and as many as possible has become the focus of attention of medical professionals and academic professionals.Before exposing to the public, clinical trials must be designed for drugs to verify their effectiveness and security. However, due to some defects, clinical trials cannot find all adverse reactions of drugs. After flowing into market, adverse drug event reporting systems become the main means of monitoring drug safety and finding potential adverse drug reactions. With the development of technologies related to Web2.0 and the popularization of the Internet, health-related social network websites have collected a huge amount of user comments about drugs from patients. These user comments contain huge amount information about adverse drug reactions. Therefore, health-related social network websites provide another data source for potential adverse drug reaction discovery.Towards the data from adverse event reporting systems, the paper trains non-sequenced Skip-gram model to generate the distributed vectors for drug entities and adverse reaction entities with these reports. With the entity vectors, the paper computes the relevancy among drugs and adverse reactions and then finds potential adverse drug reactions based on relevancy. The experiment results show that the distributed entity vectors capture the relevancy among drugs and adverse reactions effectively, and can be used to finding the potential drug reactions.Towards the user comments from social networking websites, the paper finds mentions of adverse reactions from them based on method of information entropy and dictionary matching. However, the adverse reactions found from user comments are potential, not real, adverse drug reactions because they have not gotten clinical verifications. And getting clinical verifications for potential adverse drug reactions is a process requiring much time and effort. Therefore, the paper generates distributed biomedical entity vectors using non-sequenced Skip-gram model. For drug d and adverse reaction a, the paper tries best to find some proteins, called association-proteins, that can be associated with both drug d and adverse reaction a based on the generated distributed biomedical entity vectors. The medical specialists can refer to the association-proteins when verify the reality of potential adverse drug reactions and then reduce the time of the getting clinical verifications, so that they can find the potential drug risks timely.