Paper detail

Detect adverse drug reactions for drug Atorvastatin

Adverse drug reactions (ADRs) are big concern for public health. ADRs are one of most common causes to withdraw some drugs from markets. Now two major methods for detecting ADRs are spontaneous reporting system (SRS), and prescription event monitoring (PEM). The World Health Organization (WHO) defines a signal in pharmacovigilance as "any reported information on a possible causal relationship between an adverse event and a drug, the relationship being unknown or incompletely documented previously". For spontaneous reporting systems, many machine learning methods are used to detect ADRs, such as Bayesian confidence propagation neural network (BCPNN), decision support methods, genetic algorithms, knowledge based approaches, etc. One limitation is the reporting mechanism to submit ADR reports, which has serious underreporting and is not able to accurately quantify the corresponding risk. Another limitation is hard to detect ADRs with small number of occurrences of each drug-event association in the database. In this paper we propose feature selection approach to detect ADRs from The Health Improvement Network (THIN) database. First a feature matrix, which represents the medical events for the patients before and after taking drugs, is created by linking patients' prescriptions and corresponding medical events together. Then significant features are selected based on feature selection methods, comparing the feature matrix before patients take drugs with one after patients take drugs. Finally the significant ADRs can be detected from thousands of medical events based on corresponding features. Experiments are carried out on the drug Atorvastatin. Good performance is achieved.