Automated Detection Using Natural Language Processing of Radiologists Recommendations for Additional Imaging of Incidental Findings
Automated Detection Using Natural Language Processing of Radiologists Recommendations for Additional Imaging of Incidental Findings
复制标题
DOI：
10.1016/j.annemergmed.2013.02.001
复制
发表时间：
2013-08-01
影响因子：
6.2
通讯作者：
Reisner, Andrew T.
中科院分区：
文献类型：
--
作者：
Dutta, Sayon;Long, William J.;Reisner, Andrew T.
研究目的：随着放射学检查使用的增加，偶然发现（如肺结节）也同时增加，放射科医生建议进行额外成像以进行随访。忙碌的急诊医生可能会受到挑战，需要仔细沟通与患者初步评估无关的额外成像建议。电子健康记录和自然语言处理算法的出现可能有助于解决这一质量差距。我们寻求描述的建议，从我们的机构，并开发和验证一个自动化的自然语言处理算法，以可靠地识别建议额外imaging.Methods：我们开发了一个自然语言处理算法，以检测建议额外的成像，使用3个迭代周期的训练和验证。第三个周期使用了出院急诊科（艾德）患者的3，235份放射学报告（1，600份用于算法培训，1，635份用于验证），从中我们确定了出院相关额外成像建议的发生率和适当出院记录的频率。使用盲法图表审查作为标准，比较了3种自然语言处理算法迭代的测试特征。4.5%的患者提出了出院相关的额外影像学建议（95%置信区间[CI] 3.5%至5.5%）的艾德放射学报告，但51%（95% CI 43%至59%）的出院说明没有注意到这些结果。最终的自然语言处理算法在检测额外成像建议方面具有89%（95% CI 82%至94%）的灵敏度和98%（95% CI 97%至98%）的特异性。对于出院相关的建议，额外的成像，敏感性提高到97%（95%CI 89%至100%）。结论：建议额外的成像是常见的，和失败的文件相关的建议，额外的成像在艾德出院指示经常发生。自然语言处理算法的性能随着每次迭代而提高，并提供了一个有前途的错误预防工具。
Study objective: As use of radiology studies increases, there is a concurrent increase in incidental findings (eg, lung nodules) for which the radiologist issues recommendations for additional imaging for follow-up. Busy emergency physicians may be challenged to carefully communicate recommendations for additional imaging not relevant to the patient's primary evaluation. The emergence of electronic health records and natural language processing algorithms may help address this quality gap. We seek to describe recommendations for additional imaging from our institution and develop and validate an automated natural language processing algorithm to reliably identify recommendations for additional imaging.Methods: We developed a natural language processing algorithm to detect recommendations for additional imaging, using 3 iterative cycles of training and validation. The third cycle used 3,235 radiology reports (1,600 for algorithm training and 1,635 for validation) of discharged emergency department (ED) patients from which we determined the incidence of discharge-relevant recommendations for additional imaging and the frequency of appropriate discharge documentation. The test characteristics of the 3 natural language processing algorithm iterations were compared, using blinded chart review as the criterion standard.Results: Discharge-relevant recommendations for additional imaging were found in 4.5% (95% confidence interval [Cl] 3.5% to 5.5%) of ED radiology reports, but 51% (95% Cl 43% to 59%) of discharge instructions failed to note those findings. The final natural language processing algorithm had 89% (95% Cl 82% to 94%) sensitivity and 98% (95% Cl 97% to 98%) specificity for detecting recommendations for additional imaging. For discharge-relevant recommendations for additional imaging, sensitivity improved to 97% (95% Cl 89% to 100%).Conclusion: Recommendations for additional imaging are common, and failure to document relevant recommendations for additional imaging in ED discharge instructions occurs frequently. The natural language processing algorithm's performance improved with each iteration and offers a promising error-prevention tool.