Detection of adverse events in electronic medical records using natural language processing

事件报告系统被广泛应用于医院，以发现患者安全事件和不良事件。然而，该系统存在一些问题，如报告不足和事件与提交报告之间的时间滞后。为了解决这些问题，我们试图通过使用电子医疗记录的自然语言处理来检测事件。我们构建了句法-语义类别决策规则，并将这些规则应用于医疗记录中的叙述性文本数据。对于住院病人的福尔斯，这些规则表现出令人满意的性能（灵敏度86。5%，特异性97。5%）。由于其低特异性或低敏感性，病程记录和出院总结不适合用于该方法的数据来源。相比之下，图像订单条目上的文本数据的自然语言处理使得能够快速检测有害的福尔斯。该方法可以有效地弥补事故报告系统的漏报、滞后等缺陷。

Incident reporting system is widely used in hospital to detect patient safety incidents and adverse events. However, there are several problems in the system such as under-reporting and lag-time between events and submission of the reports. To resolve these issues, we tried to detect incidents by using natural language processing of electronic medical records. We constructed syntactic-semantic category decision rules and applied these rules to narrative text data in medical records. As for inpatient falls, these rules showed satisfactory performance(sensitivity 86. 5% and specificity 97. 5%). Progress notes and discharge summaries were not appropriate for data sources to be used in this method because of their low specificity or low sensitivity. In contrast, natural language processing of text data on image order entries enables rapid detection of injurious falls. This method is useful to compensate the shortcomings of incident reporting system such as the under reporting and the lag-time.