Radiology department utilizes Facebook algorithm to anticipate future CT, MRI volumes
A free-use algorithm from Facebook can help radiology departments predict future imaging volume trends and better allocate resources to meet anticipated demand, according to evidence published Tuesday.
The Silicon Valley giant, now known as Meta, designed its open-source Prophet tool to forecast trends across multiple industries, taking events such as holidays or even weather changes into account. Memorial Sloan Kettering Cancer Center doctors applied the procedure to their radiology services, reporting promising results in the Journal of Digital Imaging.
Prophet proved “significantly” better at anticipating CT and MRI volumes compared to manual attempts. This proved particularly true during the height of the pandemic, with the algorithm coming within four scans per day of actual volumes.
“Resource planning is a critical component of success in a radiology department,” Anton S. Becker, with the New York institution’s Department of Radiology, and colleagues explained. “We found that the algorithm captures weekly, seasonal, and overall trends and allows for better radiologist allocation compared to manual planning,” the team added later.
Becker and co-investigators refined the Prophet tool using more than 610,000 exams, of which 67,180 were second reads. Another 13,961 images were used to validate the training.
After prospective testing in February 2020, the tool missed by a mean of 10 CT exams per day (9,553 actually performed vs. 9,942 forecasted). It fared better for MRI volumes, with a mean error of two per day (2,397 performed vs. 2,485 forecasted).
Memorial Sloan, like most hospitals, experienced a significant drop in imaging volumes during the pandemic’s peak (March-May 2020) along with a rebound shortly after. They tested the model in August of that year, which overshot CT usage by 10 per day, or 317 total. MRI predictions reached a mean error of four per day and 128 overall.
“The Prophet procedure was significantly more accurate than the manual forecast,” Becker et al. noted.
The team did warn the algorithm may underperform for organizations with “simple periodic” effects, such as quarterly estimates. But they do see significant opportunities for radiology and many other fields.
“We expect our results to be reproducible in other bottleneck problems in healthcare with seasonal fluctuations, such as surgical procedures or radiotherapy treatments,” they concluded. “Lastly, these methods may also prove useful in allotting resources in academic research, for example, by forecasting occupancy of laboratory animal housing facilities or eligible patients for enrollment in clinical trials.”
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.