Adverse drug events are a major cause of morbidity and mortality worldwide. In the United States alone, serious adverse events effect more than 2 million patients and cause more than 100,000 deaths every year, making them the fourth leading cause of death and disease. Methods capable of detecting and predicting these events are needed to reduce unnecessary death. For example, by the time that cerivistatin was withdrawn in 2001, 52 deaths had already occurred due to rhabdomyolysis and associated renal failure. In many cases, these adverse events could have been predicted using current biological and pharmacological knowledge.

The FDA’s adverse event database (AERS) receives more than 300,000 reports of adverse events every year. Included in each report is a list of the medications involved, the diagnoses of the patient, and the adverse events observed. For example, a report may include:

Report ID: 4223542
Drugs: Remicade, Benadryl
Diseases Diagnosed: Crohn’s Disease
Adverse Events: Arthralgia, Dyspnoea, Musculoskeletal Stiffness, Trismus

We believe an opportunity exists to combine this data with more biological knowledge to create a classifier that uses machine learning methods to predict which of these events are more likely to occur in a given clinical scenario. For example, there are databases such as the Pharmacogenomics Knowledge Base (PharmGKB) which includes relationships between genes and drugs diseases and DrugBank which includes information about drugs and their chemical structures. We believe we can combine the data from AERS with additional data from PharmGKB and DrugBank to create feature vectors which are predictive of adverse events.

Specifically, we propose to: 1) create a feature vector for each adverse event report by linking the AERS data to other biological databases that provide additional features about the drugs, indications, and adverse events involved, 2) evaluate and compare predictive methods for accuracy, training requirements, and 3) evaluate the primary predictive features to better understand the biology leading to adverse pharmacological events. These methods will provide a tool to predict and prevent the morbidity and mortality due to adverse drug events.