Mining of clinical data can protect and enhance radiology practice

Most practicing radiologists can think of an occasion when something they intuitively and absolutely knew to be true turned out, upon close and rigorous inspection, not to be true. A new tool being adopted in radiology-data mining of radiology reports, images, and overall practice patterns-promises to create many more such moments. Applied widely, this tool could dislodge a lot of old shibboleths and improve the quality of patient care.

An example of how this could happen appears in a special report in the March issue of Radiology, "Does radiologist recommendation for follow-up with the same imaging modality contribute substantially to high-cost imaging volume?" (Radiology 2007;242:857-864). In the report, researchers from Massachusetts General Hospital use data mining techniques to probe more than 2.9 million reports with a prior examination of the same type within seven months. They found that nearly one-third of high-cost imaging exams are repeat exams of the same patient. But they also found that only 8% of these high-cost exams are recommended by the radiologist in the original report.

Further, the Mass General study team found that the majority of the repeat exams are for outpatients undergoing cancer treatment and neurosurgical inpatients. They suggest, logically, that a $350 CT scan is a small price to pay to assure that a $100,000 chemotherapy regimen is working and that using repeat CT scans to monitor intracranial surgery can expedite treatment decisions and minimize hospital bed time.

At a time when high-cost imaging is under a microscope, information such as this can have a game-changing impact on the terms of the debate. Perhaps those second sets of images "self-referred" by radiologists are not all that common, and second imaging studies in general may make a lot more sense than people realize.

But the ramifications needn't end there. Another consequence of gathering this type of information is that everyone gets a better understanding of practice patterns that may or may not make sense. The Mass General study notes that oncologists, neurologists, and neurosurgeons at the facility have, as a result of the analysis, begun to review their exam-ordering practices. Their goal is to arrive at recommendations for imaging in the context of the entire clinical scenario.

In this setting, new understandings provided by data mining open the door for improvements in the quality of care. Mass General researchers note that, nationally, authors of chemotherapy regimens and algorithms for management of intracranial injury can develop guidelines for the frequency of imaging studies that consider their contribution to overall healthcare costs.

A key to making this happen is the availability of digitized radiology reports, IT skills, and a program developed at Mass General called Leximer, a computer algorithm that allows computerized searches to draw out context-sensitive findings from unstructured radiology reports.

We've seen data mining efforts in medicine in the past based on large databases from Medicare and other sources. But these have been limited generally to broad business and practice questions, such as who is billing Medicare for angiography procedures. What's different here is that the data mining has moved into the area of clinical information derived from the radiology report.

Now we'll be able to look at questions like what types of scans yield positive findings. It makes possible a closer and more targeted look at large data sets to help decide what clinical practices and settings generate the best results.

Eventually, we can anticipate that this approach will move well beyond academic settings to private practice and multi-institutional reviews. It could bring in other medical data besides radiology reports. As that happens, we can look for significant improvements in the quality of radiology and, as a consequence, medical practice.