A research team including presenter Dr. Elina Khasanova of the University of British Columbia in Canada and colleagues compared the interreader agreement of a commercial AI algorithm (HeartFlow) with three certified level 3 CCTA readers on 50 exams. The 50 CCTA exams were processed on the AI model by separate analysts.
The researchers also assessed intrarater agreement in 15 of the cases, which were interpreted three times by all readers one to two weeks apart. In addition, the same analysts processed these 15 cases three times.
The AI-based coronary stenosis assessment was more reproducible and had higher agreement than the level 3 CCTA readers, according to the researchers.
"The AI-based automated tool and with analyst input may improve readers' agreement and reproducibility in reporting of moderate stenosis and non-obstructive disease in a more targeted per-vessel and per-segment manner," the authors wrote.
They noted that the use of a reporting system with a broader stenosis "bucket" improved interreader agreement for both the readers and the algorithm. However, further larger studies are needed, including a comparison of the algorithm's diagnostic accuracy with invasive coronary angiography, according to the researchers.
Get the rest of their results at this talk on Monday afternoon.
