Biases introduced by filtering electronic health records for patients with "complete data"

项目摘要 电子健康记录（EHR）在全国范围内的采用导致了越来越多的大型临床数据的可用性。 数据集。通过统计建模和机器学习，这些数据集已被广泛用于 这些应用包括诊断、决策支持、降低成本和个性化医疗。然而，在这方面， 因为同一个病人可以在多个医疗保健机构接受治疗，所以仅来自单个EHR的数据可能 不包含该患者的完整病史，可能缺失关键事件。一个共同 解决此问题的方法是应用数据检查，对出现数据的患者的EHR进行过滤 更“完整”。过滤器的例子包括要求每年至少一次访问或确保年龄，性别， 和种族都有记录。然而，在之前的一项研究中，我们使用了来自七个机构的EHR数据，结果表明， 这些过滤器可以大大减少样本量，并通过选择病情较重的患者来引入意想不到的偏倚。 更频繁地寻求护理并改变由此产生的队列的人口统计数据。该项目扩展了这一 通过实施一套扩展的数据完整性过滤器并测试其准确性， 使用国家索赔数据和来自数十家医院的EHR数据的组合， 全国各地的医疗中心。这将使我们能够了解数据完整性如何在 不同的EHR，并量化不同的方法来纠正病人的记录差距的权衡。第一、 我们将使用国家索赔数据开发和衡量数据完整性过滤器的准确性。这提供 纵向数据的“黄金标准”，其中患者的完整病史是已知的， 他们加入保险计划的日期。在按提供程序组对数据进行分区以建模间隙之后， 在EHR数据中，我们将测试数据完整性过滤器的效果，单独和组合机器学习 模型，选择间隙较少的患者。然后，我们将测试是否过滤器引入偏见，选择病态 患者（更多的诊断，更多的访问等）或者改变他们的人口统计特征（年龄、性别和邮编 代码）。然后，我们将在EHR数据上测试过滤器，首先在一个大型医疗中心，然后在全国范围内进行测试。 由57家机构组成的网络，代表不同的地理区域、患者人群、 数据和卫生保健设施的类型。我们将通过测量过滤器是否改善了 机器学习模型用于预测住院率的性能。我们的最终目标是（a）帮助 研究人员平衡了对完整数据的需求和可能引入模型的偏差，（B）帮助 他们预测在一个EHR数据集上训练的模型在具有不同数据的其他EHR上的效果如何 完整性配置文件。

PROJECT SUMMARY Nationwide adoption of electronic health records (EHRs) has led to the increasing availability of large clinical datasets. With statistical modeling and machine learning, these datasets have been be used in a wide range of applications, including diagnosis, decision support, cost reduction, and personalized medicine. However, because the same patient could be treated at multiple health care institutions, data from only a single EHR might not contain the complete medical history for that patient, with critical events potentially missing. A common approach to addressing this problem is to apply data checks that filter the EHR for patients whose data appear to be more “complete”. Examples of filters include requiring at least one visit per year or ensuring that age, sex, and race are all recorded. However, in a previous study using EHR data from seven institutions, we showed that these filters can greatly reduce the sample size and introduce unexpected biases by selecting sicker patients who seek care more often and changing the demographics of the resulting cohorts. This project extends this prior research by implementing an expanded set of data completeness filters and testing their accuracy and potential biases using a combination of national claims data and EHR data from dozens of hospitals and healthcare centers across the country. This will enable us to understand how data completeness varies in different EHRs and quantify the tradeoffs of different approaches to correcting for gaps in patients' records. First, we will develop and measure the accuracy of data completeness filters using national claims data. This provides a “gold standard” of longitudinal data where patients' complete medical histories are known during the periods in which they were enrolled in the insurance plan. After partitioning the data by provider groups to model gaps in EHR data, we will test how well data completeness filters, individually and in combined machine learning models, select patients with fewer gaps. We will then test whether the filters introduce biases by selecting sicker patients (more diagnoses, more visits, etc.) or changing their demographic characteristics (age, sex, and zip code). Then, we will test the filters on EHR data, first at a single large medical center, and then across a national network of 57 institutions, representing different geographic regions, patient populations, number of years of data, and types of health care facilities. We will evaluate the filters by measuring whether they improve the performance of a machine learning model for predicting hospital admissions. Our ultimate goals are to (a) help researchers balance the need for complete data with the biases this might introduce to their models and (b) help them predict how well models trained on one EHR dataset might work on other EHRs with different data completeness profiles.