Just-in-time data mining addresses uncertainty in radiology reports

Tuesday, November 29 | 3:50 p.m.-4:00 p.m. | SSJ11-06 | Room S102D
Imagine sitting at a diagnostic workstation with a difficult case to interpret, and being able to retrieve similar images and related reports contained in a PACS with a single button command. This capability is being tested at Massachusetts General Hospital using prototype software.

Rad-Aid is natural language processing software that performs real-time data mining at a diagnostic workstation, enabling radiologists to compare their findings with similar reports in the PACS database.

The software communicates via an HL7 server to a RIS, radiology order-entry platform, and patient demographics. It performs an automated query of the database to identify the most common clinical findings present in the radiology report based on the patient's age, gender, exam indication, and exam modality using the image accession number as the identifier. The top 10% of the most common matches are returned to the radiologist for optional selection and review.

Imaging informatics research fellow Dr. Supriya Gupta will discuss a validation test in which 1,000 consecutive head CT reports selected by Rad-Aid were evaluated for accuracy. Researchers found that 81% of the reports contained similar patient demographics and clinical indications. The highest concordance rate occurred when the report identified a mass, metastatic lesion, edema, or reported ischemic changes. More specific findings such as calcification, chondrosarcoma, cyst, meningioma, and osteoma had the lowest correlation.

Page 1 of 603
Next Page