Science Applications International Corporation, USA.
Int J Med Inform. 2011 Jan;80(1):67-73. doi: 10.1016/j.ijmedinf.2010.10.013. Epub 2010 Nov 19.
Near real-time disease detection using electronic data sources is a public health priority. Detecting pneumonia is particularly important because it is the manifesting disease of several bioterrorism agents as well as a complication of influenza, including avian and novel H1N1 strains. Text radiology reports are available earlier than physician diagnoses and so could be integral to rapid detection of pneumonia. We performed a pilot study to determine which keywords present in text radiology reports are most highly associated with pneumonia diagnosis.
Electronic radiology text reports from 11 hospitals from February 1, 2006 through December 31, 2007 were used. We created a computerized algorithm that searched for selected keywords ("airspace disease", "consolidation", "density", "infiltrate", "opacity", and "pneumonia"), differentiated between clinical history and radiographic findings, and accounted for negations and double negations; this algorithm was tested on a sample of 350 radiology reports. We used the algorithm to study 189,246 chest radiographs, searching for the keywords and determining their association with a final International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis of pneumonia.
Performance of the search algorithm in finding keywords, and association of the keywords with a pneumonia diagnosis.
In the sample of 350 radiographs, the search algorithm was highly successful in identifying the selected keywords (sensitivity 98.5%, specificity 100%). Analysis of the 189,246 radiographs showed that the keyword "pneumonia" was the strongest predictor of an ICD-9-CM diagnosis of pneumonia (adjusted odds ratio 11.8) while "density" was the weakest (adjusted odds ratio 1.5). In general, the most highly associated keyword present in the report, regardless of whether a less highly associated keyword was also present, was the best predictor of a diagnosis of pneumonia.
Empirical methods may assist in finding radiology report keywords that are most highly predictive of a pneumonia diagnosis.
利用电子数据源进行近乎实时的疾病检测是公共卫生的当务之急。检测肺炎尤为重要，因为它是几种生物恐怖主义制剂的表现疾病，也是流感（包括禽流感和新型 H1N1 病毒）的并发症。放射学文本报告比医生诊断更早可用，因此对于快速检测肺炎可能是不可或缺的。我们进行了一项试点研究，以确定文本放射学报告中哪些出现的关键词与肺炎诊断高度相关。
使用了来自 11 家医院的 2006 年 2 月 1 日至 2007 年 12 月 31 日的电子放射学文本报告。我们创建了一个计算机算法，该算法可搜索选定的关键词（“气腔疾病”、“实变”、“密度”、“浸润”、“不透明”和“肺炎”），区分临床病史和放射学发现，并考虑到否定和双重否定；该算法在 350 份放射学报告的样本上进行了测试。我们使用该算法研究了 189246 张胸片，搜索这些关键词并确定它们与国际疾病分类第 9 版临床修订版（ICD-9-CM）肺炎诊断的关联。
搜索算法查找关键词的性能，以及关键词与肺炎诊断的关联。
在 350 份放射照片样本中，搜索算法在识别选定的关键词方面非常成功（灵敏度 98.5%，特异性 100%）。对 189246 张 X 光片的分析表明，关键词“肺炎”是 ICD-9-CM 肺炎诊断的最强预测因素（调整后的优势比 11.8），而“密度”是最弱的预测因素（调整后的优势比 1.5）。一般来说，无论报告中是否存在关联性较弱的关键词，报告中出现的关联性最强的关键词都是肺炎诊断的最佳预测因素。
实证方法可能有助于找到与肺炎诊断高度相关的放射学报告关键词。
