Dang Pragya A, Kalra Mannudeep K, Blake Michael A, Schultz Thomas J, Halpern Elkan F, Dreyer Keith J
Department of Radiology, Massachusetts General Hospital, 25 New Chardon St., Ste. 400E, Boston, MA 02114, USA.
AJR Am J Roentgenol. 2008 Aug;191(2):313-20. doi: 10.2214/AJR.07.3508.
The purposes of this study were to validate a natural language processing program for extraction of recommendation features, such as recommended time frames and imaging technique, from electronic radiology reports and to assess patterns of recommendation features in a large database of radiology reports.
This study was performed on a radiology reports database covering the years 1995-2004. From this database, 120 reports with and without recommendations were selected and randomized. Two radiologists independently classified these reports according to presence of recommendations, time frame, and imaging technique suggested for follow-up or repeated examinations. The natural language processing program then was used to classify the reports according to the same criteria used by the radiologists. The accuracy of classification of recommendation features was determined. The program then was used to determine the patterns of recommendation features for different patients and imaging features in the entire database of 4,211,503 reports.
The natural language processing program had an accuracy of 93.2% (82/88) for identifying the imaging technique recommended by the radiologists for further evaluation. Categorization of recommended time frames in the reports with the 88 recommendations obtained with the program resulted in 83 (94.3%) accurate classifications and five (5.7%) inaccurate classifications. Recommendations of CT were most common (27.9%, 105,076 of 376,918 reports) followed by those for MRI (17.8%). In most (85.4%, 322,074/376,918) of the reports with imaging recommendations, however, radiologists did not specify the time frame.
Accurate determination of recommended imaging techniques and time frames in a large database of radiology reports is possible with a natural language processing program. Most imaging recommendations are for high-cost but more accurate radiologic studies.
本研究的目的是验证一个自然语言处理程序，用于从电子放射学报告中提取推荐特征，如推荐的时间框架和成像技术，并评估大型放射学报告数据库中推荐特征的模式。
本研究在一个涵盖1995年至2004年的放射学报告数据库上进行。从该数据库中，选择了120份有和没有推荐意见的报告并进行随机分组。两名放射科医生根据推荐意见的存在、时间框架以及为后续或重复检查建议的成像技术，对这些报告进行独立分类。然后使用自然语言处理程序根据放射科医生使用的相同标准对报告进行分类。确定推荐特征分类的准确性。然后使用该程序确定整个4,211,503份报告数据库中不同患者和成像特征的推荐特征模式。
自然语言处理程序在识别放射科医生推荐用于进一步评估的成像技术方面的准确率为93.2%（82/88）。使用该程序获得的88条推荐意见的报告中，推荐时间框架的分类结果为83条（94.3%）准确分类和5条（5.7%）不准确分类。CT的推荐最为常见（27.9%，376,918份报告中的105,076份），其次是MRI（17.8%）。然而，在大多数（85.4%，322,074/376,918）有成像推荐意见的报告中，放射科医生没有指定时间框架。
使用自然语言处理程序可以在大型放射学报告数据库中准确确定推荐的成像技术和时间框架。大多数成像推荐是针对高成本但更准确的放射学检查。
