Program uses EHR to flag VTE cases in narrative radiology reports

An EHR search could help hospitals improve how they spot venous thromboembolism (VTE) cases. Using a bag-of-words approach, a model accurately and swiftly identified deep vein thrombosis (DVT) and pulmonary embolism (PE) cases from radiology reports.

A McGill University research team in Montreal and colleagues developed an automated model to identify cases of VTE and improve accuracy of a computer model without resorting to a more labor intensive manual method. The research team noted that identifying these patients quickly and accurately from the radiology reports can better help focus hospital improvement efforts to reduce adverse events. They considered development of a computer model to be a first step in creating a near-real time adverse event monitoring approach scalable to entire hospital systems.

The model was built using data extracted from McGill University Health Center databases and the initial analysis spanned 2008 through 2012. The initial sample included 2,000 patients suspected of DVT or PE. Two sets of models were developed, one to search for DVT and the other for PE.  Both were initially manually coded to create a reference standard and a matrix was created from the list of terms developed through initial coding. The word lists were improved on to include word combinations.

The programs successfully identified DVT in 80 percent of radiology reports, with a false positive rate of about 11 percent. PE was successfully identified in 79 percent of cases, with 16 percent false positive rate. These rates were an improvement over those noted in prior research in this area.

“Overall, the results of our study add to this emerging body of literature and provide further evidence that automated methods based on NLP [natural language processing] techniques can successfully be applied to EHR data for the purpose of identifying adverse events such as DVT and PE,” wrote Christian M. Rochefort, RN, PhD, of the Clinical and Health Informatics Research Group at McGill, and colleagues.  

VTE, they noted, is seen in up to 40 percent of medical and general surgical patients, with even higher rates in patients undergoing major orthopedic procedures. Current estimates on 30-day fatality rates in patients who develop VTEs range from 5 to 15 percent and it is the second most common cause of increased hospital stay lengths. Rochefort et al noted that focusing on this subset of adverse events would significantly improve patient outcomes.

They proposed this approach also could be used to detect other types of adverse events.

Rochefort et al suggested that future improvements to modeling approaches might include both a syntactical and word-based approach. “The combination of symbolic and statistical NLP techniques represents an interesting area of future research and promises improved performance in adverse event detection,” they wrote.

This study was published online Oct. 20 in the Journal of the American Medical Informatics Association.