Publicly available AI algorithm proves useful for spotting actionable radiology findings

PARCA eNews – Sept. 12, 2021 – A publicly available artificial intelligence algorithm for detecting and flagging "actionable" radiology reports proved superior to three other methods for distinguishing actionable from non-actionable reports.

The study by researchers let by Yuta Nakamura at the University of Tokyo appears in the Sept. 11 journal, BMC Medical informatics and Decision Making. The goal is to use AI to more rapidly detect and refer radiological findings to clinicians.
“The results of this study suggest that the radiologists may have sometimes thought that actionable findings were present in the radiological images without explicitly urging further clinical examinations or treatments in the radiology report,” Nakamura et al. wrote, adding their method can key in on these cases and bring them to the physicians’ attention.

In the study the researchers used 90,923 confirmed radiology reports for computed tomography examinations performed by the university’s radiology department between Sept. 9, 2019 and April 30, 2021. The reports covered several areas of the body including brain, head and neck, cardiac, and musculoskeletal body parts.

The reports were all tagged as actionable or non-actionable by the diagnosing radiologists, which served as the gold standard that the four algorithms would be judged. Overall there were 788 reports tagged as actionable.

Each of the four machine-learning algorithms "read" all 90,923 reports separating the actionable from the non-actionable. In the end the publicly available one that used bidirectional encoder representations from transformers (BERT) proved superior to the others.

The researchers state in their paper that BERT may have performed better for two reasons: 1) BERT can use linguistic knowledge from both an in-house dataset and outside set of documents, or corpus. 2) BERT is able to capture the relationship between two documents.

“The results showed that our method based on BERT is more useful for distinguishing various actionable radiology reports from non-actionable ones than models based on other deep learning methods or statistical machine learning,” the group wrote.

Source: BMC Medical Informatics and Decision Making