Automatic structuring of radiology reports: harbinger of a second information revolution in radiology.

Over the past 2 decades, rapid technologic advances have enabled radiologists to acquire, transmit, process, display, and store multidimensional digital images. Despite the steady advances in the way radiologists manage images, similar technologic innovations have not yet affected how radiologists communicate with their colleagues. This discrepancy is particularly striking because the information in radiology reports is just as amenable to computer-based storage, processing, and display as pixels, voxels, and images. In this issue of the journal, Dr Hripcsak and colleagues (1) provide a welcome early look at how information technology will shape the future of radiology reporting (1). They have developed a system that uses natural language processing (NLP) to dissect and structure meaty clinical prose into small digital morsels, each containing a unique medical concept suitable for processing by a computer. Because conventional narrative radiology reports served as early stimuli to refine and improve their NLP system, rigorous evaluations have shown that its accuracy in extracting information from radiology reports is comparable with that of human experts. Their report details an ambitious and impressive evaluation of the NLP system on a huge database of clinical radiology reports encompassing a quarter of a million patients. The authors tasked their system with creating a semantic structure for nearly 900,000 chest radiography reports, which were dictated by radiologists at Columbia-Presbyterian Medical Center over a decade of patient care. An ingenious method was used to externally validate the contents of the structured reports. The frequency and co-occurrence of a variety of clinical conditions was computed from the structured reports and compared with preselected external benchmarks, including peer-reviewed medical literature, regional crime statistics, and financial coding results. These comparisons demonstrate that the structured reports accurately reflected the 3:2 right-to-left ratio of lung cancer, the association of pleural effusions with other clinical conditions, and the decreasing incidence of violent crime in their catchment area. Although not externally validated, their study also produced a fascinating portrait of the semantic contents of radiology reports. For example, their Table 1 provides a unique descriptive look at the occurrence of a variety of clinical conditions in chest radiography reports that ranged from tuberculosis to rib fracture. A separate analysis quantified the radiologist’s impression that a substantial fraction of reports are normal—30% of chest radiography reports in their analysis. The accuracy was comparable with that of expert human coders—achieving 81% sensitivity and 99% specificity, confirming the results of Hripcsak and colleagues’ (1) previous studies. Furthermore, the authors showed that their software outperformed human financial discharge coders in recognizing certain important clinical conditions, such as pneumothorax. Taken together, their results provide a rare look at the contents of radiology reports and illustrate the feasibility of NLP methods to accurately transform databases containing narrative radiology reports into rich sources of information for radiology practices.