Veterans Affairs researchers are using natural language processing technologies to flag veterans' electronic health records for evidence of a predilection toward suicide.
In 2012 more American soldiers died by suicide than in direct combat. A year later, health care and IT researchers working with the Veterans Affairs Department are looking to use natural language processing (NLP) to mine electronic health record (EHR) systems for telltale signs of soldiers at clear risk for suicide.
“The electronic medical record system stores a very large body of clinical notes,” Dr. Ken Hammond, a retired VA psychiatrist who is helping with the research, told the VA news service. In fact, VA says it securely stores EHRs on 14 million current and former patients. The records contain 2 billion documents in all.
By using new NLP and search techniques, the VA hopes to be able to flag patients who present clear risk of suicide. “We’ve shown that we can use search engine technology to more easily identify those veterans who have attempted suicide at some point in their lives,” Hammond said. “That can help us prevent future attempts.”
Hammond headed a group that identified search terms to query doctors’ free-text notes stored in the EHRs of more than 100,000 veterans. The tools were designed to use NLP techniques to highlight updates in the records indicating a veteran may have had a past suicide attempt.
Free text is more difficult to crunch than check boxes and other structured data, but researchers are using the technology to program the computers to spot meaningful phrases, according to the VA news service.
Agencies from NASA to the Homeland Security Department are using natural language processing for text analytics projects ranging from scanning social media to checking airline logs for safety warnings.
According to researchers, a big challenge was distinguishing between a remark or phrase about suicide in the screening notes and an apparent actual suicide attempt, the VA service reported. In this case, researchers wanted to highlight documentation that did indicate a suicide attempt by the patient.
In doing so, they developed an automated text search that was about 80 percent accurate, compared with a doctor manually checking each record.
Hammond, together with VA colleagues in Salt Lake City and Boston, presented the results of their research earlier this year at the Hawaii International Conference on System Sciences.