Accounting for Differential Error in Time-to-Event Analyses Using Imperfect Electronic Health Record-Derived Endpoints
Accounting for Differential Error in Time-to-Event Analyses Using Imperfect Electronic Health Record-Derived Endpoints
复制标题
DOI：
10.1007/978-3-319-69416-0_14
复制
发表时间：
2017-01-01
期刊：
影响因子：
--
通讯作者：
Chubak, Jessica
中科院分区：
文献类型：
--
作者：
Hubbard, Rebecca A.;Harton, Joanna;Chubak, Jessica
由于对感兴趣的结果的不完全确定，对结果和暴露之间的关系的估计是有偏差的。在使用来自电子健康记录(EHR)的数据的研究中，结果的错误分类很常见，而且往往与患者的特征有关。例如，并存疾病负担较大的患者可能会更频繁地使用医疗系统，使EHR更有可能包含他们的诊断记录，可能导致不经常寻求治疗的较健康患者的预后分类较差。这在对事件发生时间结果的研究中尤其有问题，在该研究中，事件的发生和事件的时间(如果发生)都可能在电子病历中错误地捕捉到。使用离散时间比例风险模型可以获得在事件间隔时间结果方面的错误分类调整的估计值，但如果EHR衍生的终点的操作特征在暴露类别中不同，则可能是有偏差的。受一种使用EHR数据识别第二次乳腺癌事件的算法的启发，我们研究了在事件间隔时间分析中使用具有差异测量误差的不完全评估结果的含义。我们使用模拟研究来证明由于未能解释复发状态和时间的错误而导致的偏差的大小，并比较了纠正这种偏差的替代方法。我们总结了使用电子病历数据对事件间隔时间研究中结果错误分类进行解释的一般指南。
Estimates of the relationship between an outcome and an exposure are biased by imperfect ascertainment of the outcome of interest. In studies using data derived from electronic health records (EHRs), misclassification of outcomes is common and is often related to patient characteristics. For instance, patients with greater comorbid disease burden may use the healthcare system more frequently making it more likely that the EHR will contain a record of their diagnosis, possibly resulting in poorer outcome classification for healthier patients who do not seek care as frequently. This is particularly problematic in studies of time-to-event outcomes in which both the occurrence of an event and the timing of the event, if it occurs, may be captured with error in the EHR. Misclassification-adjusted estimators in the context of time-to-event outcomes are available using discrete time proportional hazards models but may be biased if operating characteristics of the EHR-derived endpoint vary across exposure categories. Motivated by an algorithm for identifying second breast cancer events using EHR data, we investigated the implications of using an imperfectly assessed outcome with differential measurement error in time-to-event analyses. We used simulation studies to demonstrate the magnitude of bias induced by failure to account for error in the status and timing of recurrence and compared alternative methods for correcting this bias. We conclude with general guidance on accounting for outcome misclassification in time-to-event studies using EHR data.