Walker Alexander M, Zhou Xiaofeng, Ananthakrishnan Ashwin N, Weiss Lisa S, Shen Rongjun, Sobel Rachel E, Bate Andrew, Reynolds Robert F
WHISCON, Newton, MA 02466, USA.
Epidemiology, Research and Development, Worldwide Safety and Regulatory, Pfizer, New York, NY 10017, USA.
Int J Med Inform. 2016 Feb;86:62-70. doi: 10.1016/j.ijmedinf.2015.10.005. Epub 2015 Oct 31.
To describe how computer-assisted presentation of case data can lead experts to infer machine-implementable rules for case definition in electronic health records. As an illustration the technique has been applied to obtain a definition of acute liver dysfunction (ALD) in persons with inflammatory bowel disease (IBD).
The technique consists of repeatedly sampling new batches of case candidates from an enriched pool of persons meeting presumed minimal inclusion criteria, classifying the candidates by a machine-implementable candidate rule and by a human expert, and then updating the rule so that it captures new distinctions introduced by the expert. Iteration continues until an update results in an acceptably small number of changes to form a final case definition.
The technique was applied to structured data and terms derived by natural language processing from text records in 29,336 adults with IBD. Over three rounds the technique led to rules with increasing predictive value, as the experts identified exceptions, and increasing sensitivity, as the experts identified missing inclusion criteria. In the final rule inclusion and exclusion terms were often keyed to an ALD onset date. When compared against clinical review in an independent test round, the derived final case definition had a sensitivity of 92% and a positive predictive value of 79%.
An iterative technique of machine-supported expert review can yield a case definition that accommodates available data, incorporates pre-existing medical knowledge, is transparent and is open to continuous improvement. The expert updates to rules may be informative in themselves. In this limited setting, the final case definition for ALD performed better than previous, published attempts using expert definitions.
描述计算机辅助呈现病例数据如何引导专家推断电子健康记录中病例定义的机器可实现规则。作为示例，该技术已应用于获取炎症性肠病（IBD）患者急性肝功能障碍（ALD）的定义。
该技术包括从满足假定最小纳入标准的丰富人群库中反复抽取新的一批病例候选者，通过机器可实现的候选规则和人类专家对候选者进行分类，然后更新规则，使其捕捉专家引入的新差异。迭代持续进行，直到更新导致形成最终病例定义所需的更改数量可接受地少。
该技术应用于29336例IBD成年患者的结构化数据以及通过自然语言处理从文本记录中得出的术语。在三轮过程中，随着专家识别出例外情况，该技术产生的规则预测价值不断提高；随着专家识别出缺失的纳入标准，规则的敏感性也不断提高。在最终规则中，纳入和排除术语通常与ALD发病日期相关。在独立测试轮中与临床审查进行比较时，得出的最终病例定义敏感性为92%，阳性预测值为79%。
机器支持的专家审查迭代技术可以产生一个适应可用数据、纳入现有医学知识、透明且易于持续改进的病例定义。专家对规则的更新本身可能具有参考价值。在这种有限的情况下，ALD的最终病例定义比之前发表的使用专家定义的尝试表现更好。
