OR WAIT null SECS
An AI tool identified a near five-fold higher amount of potential fractures than the manual method.
Jacqueline Center, MBBS, FRACP, PhD
An artificial intelligence (AI) model outperformed a traditional method in reading reports and flagging patients with broken bones at risk of osteoporosis, according to the findings of new research to be presented at the Endocrine Society (ENDO) 2020 Annual Scientific Sessions.
Jacqueline Center, MBBS, FRACP, PhD, and a team of Australia-based investigators aimed to address the treatment gap for patients with osteoporosis who fracture and are not treated. The team developed the X-Ray Artificial Intelligence Tool software (XRAIT), which identified a near five-fold higher number of potential fractures than the manual method.
“By improving identification of patients needing osteoporosis treatment or prevention, XRAIT may help reduce the risk of a second fracture ad the overall burden of illness and death from osteoporosis," Center, from the Garvan Institute of Medical Research in Australia, said in a statement.
Center and colleagues developed XRAIT to speed up the process of reading radiology reports because manually reading the records could lead to missing at-risk people or detecting them too slowly. The AI software used natural language processing to perform a search of reports for fracture and related terms. XRAIT could be trained for site-specific reporting styles and used rule to refine identification.
XRAIT searched 5,089 digital radiology reports from patients >50 years old who went to a hospital’s emergency department and had bone imagine over a three-month period.
The team then compared the results with the manual review of the records of 224 patients referred to the hospital’s fracture liaison service during the same time. XRAIT detected 349 people with fractures, likely because of low bone mass, compared with 98 patients identified by the manual method—a more than three-fold higher detection rate.
Center and the investigators then tested XRAIT on the digitized radiology reports of a population of Australian adults >60 years old from the Dubbo Osteoporosis Epidemiology Study. Of 327 reports of confirmed known fractures and nonfractures, the AI accurately identified fractures nearly 7 out of 10 times. What’s more, the software correctly screened out those without fractures more than 9 of 10 times.
The model, unadjusted for the local radiology reporting styles in the Dubbo Osteoporosis Epidemiology Study, had a sensitivity of 69.6% and a specificity of 95%.
Approximately 44 million Americans are at risk of osteoporosis and have an increased risk of fractures because of low bone mass. Currently, only 2 in 10 older women in the US who have a fracture get treatment or tested for osteoporosis.
Traditionally, hospitals have fracture liaison services which identify patients whose fracture may be due to osteoporosis. But the AI software can speed up the process and detect patients who might have been missed during a manual reading of the radiology records. Due to limited healthcare resources, XRAIT can be optimized to manage patients identified as at-risk instead of on the identification process.
The study, “Natural Language Processing of Radiology Reports Improves Identification of Patients with Fracture,” will be presented at the Endocrine Society (ENDO) 2020 Annual Scientific Sessions.