Electronic health records (EHRs) are an increasingly utilized resource for clinical research. While their size allows for many analytical opportunities, as with most observational data there is also the potential for bias. One of the key sources of bias in EHRs is what we term informed presence-the notion that inclusion in an EHR is not random but rather indicates that the subject is ill, making people in EHRs systematically different from those not in EHRs. In this article, we use simulated and empirical data to illustrate the conditions under which such bias can arise and how conditioning on the number of health-care encounters can be one way to remove this bias. In doing so, we also show when such an approach can impart M bias, or bias from conditioning on a collider. Finally, we explore the conditions under which number of medical encounters can serve as a proxy for general health. We apply these methods to an EHR data set from a university medical center covering the years 2007-2013.

Objective: Prior studies have looked at NEWS performance in predicting in-hospital deterioration and death, but data are lacking with respect to patient outcomes following implementation of National Early Warning Score (NEWS). We sought to determine the effectiveness of NEWS implementation on predicting and preventing patient deterioration in a clinical setting. Design: Retrospective cohort study Setting: Tertiary care academic facility and a community hospital. Patients: Patients 18 years of age or older hospitalized from March 1, 2014 to February 28, 2015 during pre-implementation of NEWS to August 1, 2015 to July 31, 2016 after NEWS was implemented. Intervention(s): Implementation of NEWS within the electronic health record (EHR) and associated best practice alert. Measurements and Main Results: In this study of 85,322 patients (42,402 patients pre-NEWS and 42,920 patients post-NEWS implementation) the primary outcome of rate of ICU transfer or death did not change after NEWS implementation, with adjusted HRs of 0.94 (0.84, 1.05) and 0.90 (0.77, 1.05) at our academic and community hospital respectively. In total, 175,357 BPAs fired during the study period, with the BPA performing better at the community hospital than the academic and predicting an event within 12 hours 7.4% versus 2.2% of the time, respectively. Re-training NEWS with newly generated hospital-specific coefficients improved model performance. Conclusions: At both our academic and community hospital, NEWS had poor performance characteristics and was generally ignored by frontline nursing staff. As a result, NEWS implementation had no appreciable impact on defined clinical outcomes. Refitting of the model using site specific data improved performance and supports validating predictive models on local data.

BackgroundChronic kidney disease (CKD) is an adverse prognostic marker for valve intervention patients; however, the prevalence and related outcomes of valvular heart disease in CKD patients is unknown.Methods and ResultsIncluded patients underwent echocardiography (1999–2013), had serum creatinine values within 6 months before index echocardiogram, and had no history of valve surgery. CKD was defined as diagnosis based on the International Classification of Diseases, Ninth Revision or an estimated glomerular filtration rate <60 mL/min per 1.73 m2. Qualitative assessment determined left heart stenotic and regurgitant valve lesions. Cox models assessed CKD and aortic stenosis (AS) interaction for subsequent mortality; analyses were repeated for mitral regurgitation (MR). Among 78 059 patients, 23 727 (30%) had CKD; of these, 1326 were on hemodialysis. CKD patients were older; female; had a higher prevalence of hypertension, hyperlipidemia, diabetes, history of coronary artery bypass grafting/percutaneous coronary intervention, atrial fibrillation, and heart failure ≥mild AS; and ≥mild MR (all P<0.001). Five‐year survival estimates of mild, moderate, and severe AS for CKD patients were 40%, 34%, and 42%, respectively, and 69%, 54%, and 67% for non‐CKD patients. Five‐year survival estimates of mild, moderate, and severe MR for CKD patients were 51%, 38%, and 37%, respectively, and 75%, 66%, and 65% for non‐CKD patients. Significant interaction occurred among CKD, AS/MR severity, and mortality in adjusted analyses; the CKD hazard ratio increased from 1.8 (non‐AS patients) to 2.0 (severe AS) and from 1.7 (non‐MR patients) to 2.6 (severe MR).ConclusionsPrevalence of at least mild AS and MR is substantially higher and is associated with significantly lower survival among patients with versus without CKD. There is significant interaction among CKD, AS/MR severity, and mortality, with increasingly worse outcomes for CKD patients with increasing AS/MR severity.

Electronic health record (EHR) data are becoming a primary resource for clinical research. Compared to traditional research data, such as those from clinical trials and epidemiologic cohorts, EHR data have a number of appealing characteristics. However, because they do not have mechanisms set in place to ensure that the appropriate data are collected, they also pose a number of analytic challenges. In this paper, we illustrate that how a patient interacts with a health system influences which data are recorded in the EHR. These interactions are typically informative, potentially resulting in bias. We term the overall set of induced biases informed presence. To illustrate this, we use examples from EHR based analyses. Specifically, we show that: 1) Where a patient receives services within a health facility can induce selection bias; 2) Which health system a patient chooses for an encounter can result in information bias; and 3) Referral encounters can create an admixture bias. While often times addressing these biases can be straightforward, it is important to understand how they are induced in any EHR based analysis.