Mining Complex Clinical Data for Patient Safety Research

医疗事故伤害病人，耗费金钱，并破坏卫生保健系统。减少错误的第一步是发现错误，因为无法发现的错误就无法管理。许多方法已应用于医疗错误检测，包括强制性事件报告、自愿未遂报告、图表审查和使用信息系统的自动监视。自动化监控保证了大规模的检测，最少的人工，并且有可能实时检测以防止错误或从错误中恢复。不幸的是，大量重要的临床信息被锁定在叙述性报告中，无法用于自动决策支持系统。医学信息学和计算机科学中出现了许多工具——自然语言处理、可视化工具和机器学习——以及理解认知过程的方法。我们假设电子病历包含了对检测错误有用的信息，而自然语言处理和其他工具将允许我们检索这些信息。我们将组建一支擅长自然语言处理、数据挖掘、术语、患者安全研究和医疗保健的团队。我们将使用一个拥有200万患者10年数据的临床资料库。它包括行政、实验室和药房编码信息，以及广泛的叙述性报告，包括出院摘要、手术报告、门诊记录、尸检报告、住院登记记录、护理记录，以及来自众多辅助服务（放射学、病理学等）的报告。我们将使用经过验证的自然语言处理器MedLEE对信息进行编码，并测量自动查询的准确性，以检测和描述错误。我们将针对以下几个方面：医疗记录中的明确错误报告、NYPORTS强制事件报告、记录中的临床冲突以及其他错误信息来源。我们将使用错误和认知分析的系统方法来发现线索，以改进错误检测。我们将把该系统纳入医院当前的事件监控程序，并评估对错误检测的影响。我们将遵守严格的隐私政策和安全程序。该项目提供了一个独特的机会，将最先进的医学语言处理系统应用于大型综合临床知识库，以推进患者安全研究。

Medical errors hurt patients, cost money, and undermine the health care system. The first step to reducing errors is detecting them, for what cannot be detected cannot be managed. A number of approaches have been applied to medical error detection, including mandatory event reporting, voluntary near-miss reporting, chart review, and automated surveillance using information systems. Automated surveillance promises large-scale detection, minimal labor, and, potentially, detection in real time to prevent or recover from errors. Unfortunately, large amounts of important clinical information lie locked in narrative reports, unavailable to automated decision support systems. A number of tools have emerged from medical informatics and computer science- natural language processing, visualization tools, and machine learning- as well as methods for understanding cognitive processes. We hypothesize that the electronic medical record contains information useful for detecting errors and that natural language processing and other tools will allow us to retrieve the information. We will assemble a team skilled in natural language processing, data mining, terminology, patient safety research, and health care. We will use a clinical repository with ten years of data on two million patients. It includes administrative, laboratory, and pharmacy coded information as well as a wide range of narrative reports including discharge summaries, operative reports, outpatient notes, autopsy reports, resident signout notes, nursing notes, and reports from numerous ancillary services (radiology, pathology, etc.). We will apply a proven natural language processor called MedLEE to code the information and measure the accuracy of automated queries to detect and characterize errors. We will target several areas: explicit error reporting in the medical record, NYPORTS mandatory event reporting, clinical conflicts in record, and other sources of error information. We will use a systems approach to errors and cognitive analysis to uncover cues to improve error detection. We will incorporate the system into the hospital's current event surveillance program and assess the impact on error detection. We will adhere to strict privacy policies and security procedures. This project represents a unique opportunity to apply the most advanced medical language processing system to a large, comprehensive clinical repository to advance patient safety research.