Semantic Relatedness for Active Medication Safety and Outcomes Surveillance
			Grant 1R01LM009623-01A2 
Awarded by the National Institutes of Health, National Library of Medicine

	$935,000 over 4 years, September 2008 - September 2012

			Serguei Pakhomov (PI) 
			Ted Pedersen (co-PI)

ABSTRACT :
----------

Medication-related morbidity and mortality in ambulatory care in the 
United States results in estimated 100,000 deaths and $177 billion 
spending annually. Post-marketing passive surveillance of outcomes 
associated with medication use has been recognized as a necessary 
component in drug safety monitoring to overcome the limitations of pre- 
marketing clinical trials. Information technology applied to the 
patient's electronic medical and therapeutic record holds promise to 
improve this situation by detecting alarming trends in signs and 
symptoms in patient populations exposed to the same medication. 

Currently, much of the information necessary for active drug safety 
surveillance is "locked" in the unstructured text of electronic 
records. Our long-term goal is to develop information technology to 
recognize and prevent drug therapy related adverse events. 
Sophisticated natural language processing systems have been developed 
to find medical terms and their synonyms in the unstructured text and 
use them to retrieve information. In order to monitor alarming trends 
in symptoms in medical records, we need mechanisms that will allow not 
only accurate term and concept identification but also grouping of 
semantically related concepts that may not necessarily be synonymous. 
Measures of semantic relatedness rely on existing ontologies of domain 
knowledge as well as large textual corpora to compute a numeric score 
indicating the strength of relatedness between two concepts. Our 
central hypothesis is that such measures will be able to make 
fine-grained distinctions among concepts in the biomedical text, and 
provide a foundation upon which to organize concepts into meaningful 
groups automatically. 

In particular, this proposal seeks to develop methods that leverage the 
medical knowledge contained within Unified Medical Language System 
(UMLS) and corpora of clinical text. Our short-term goals are 1) 
develop new methods, specific to clinical text, for computing semantic 
relatedness 2) integrate these specific methods for computing semantic 
relatedness into more general methods of natural language processing 3) 
integrate semantic relatedness into methods for identifying labeled 
semantic relations in clinical text. Labeled relations significantly 
enhance the ability of natural language processing to support accurate 
automatic analysis of medical information for improving patient safety. 
Our next step will be to develop and validate a generalizable active 
medication safety surveillance system that will automatically track 
medication exposure and alarming trends in signs and symptoms in 
ambulatory and hospitalized populations for a broad range of diseases.

PUBLIC HEALTH RELEVANCE: 
------------------------

This project will a) create and validate a common open-source platform 
for developing and testing semantic relatedness measures, b) determine 
the validity of electronic medical records with respect to 
identification of symptoms associated with medication- related problems 
and c) develop a novel methodology to aggregate adverse reaction terms 
used to code spontaneous post-marketing drug safety surveillance 
reports. The results of this project will enable more effective 
medication safety surveillance efforts and thus will improve patient 
safety.