A set of newly developed algorithms can help providers identify venous thromboembolisms from radiology reports, a potentially useful tool for bolstering quality care.
That’s according to a new multi-institutional study comparing the accuracy of using ICD-10 codes and natural language processing to spot VTE in hospitalized patients. The algorithms were particularly good at finding deep venous thrombosis (DVT) within ultrasound reports and at spotting pulmonary embolism (PE) on CT reports, the authors noted in the January 2022 issue of Thrombosis Research.
Both PE and DVT are major causes of morbidity and mortality, costing U.S. healthcare providers between $15-$34 billion each year. Given these stats, the researchers decided to make their algorithms freely available to the public.
“Improving the ability to accurately identify VTE from hospital clinical and administrative databases may have important implications for research and quality improvement,” Amol A. Verma, MD, with the University of Toronto’s Department of Medicine and co-authors wrote in the study.
The team trained four tools on a set of 1,551 general internal medicine hospitalizations recorded across five Toronto hospitals between April 2010 and March 2017. Each algorithm was developed to spot PE and DVT from radiologist reports of thoracic CT, extremity compression ultrasound and nuclear ventilation-perfusion scans.
Overall, the new algorithms beat out ICD-10 codes and a previously published “simpleNLP” tool at identifying DVT and PE.
To date, administrative diagnostic codes have largely guided institutional understanding of VTE-related healthcare utilization, disease history, complications and other factors. But these findings suggest such codes are “unreliable” in spotting thromboembolism.
Verma and co-authors did acknowledge radiology reporting practices may vary across cities, countries or hospital types, limiting the results of their study. Despite this, and other potential pitfalls, the team sees a ripe opportunity for others to use their tools.
“Many large administrative databases do not presently contain radiology reports. However, electronic health records are increasingly being used for research, enabling widespread use of NLP to extract information from radiology reports,” the authors added. “Thus, we have freely shared our NLP algorithms in the hopes that they will be of value to the academic community.”
Read the entire study here.
Matt joined Chicago’s TriMed team in 2018 covering all areas of health imaging after two years reporting on the hospital field. He holds a bachelor’s in English from UIC, and enjoys a good cup of coffee and an interesting documentary.