Aernoud Fiolet

347 Accuracy of using routinely collected electronic healthcare data to identify cardiovascular endpoints DISCUSSION This analysis of the data of a randomized clinical outcome trial investigated the accuracy of endpoint identification using routinely collected EHR data as compared to conventional investigator-reported endpoints by patient interviews. The most notable findings include the following: First, we found that almost all trial participants could be identified using an EHR search algorithm. Second, we found high sensitivity and negative predictive values for this method to detect the composite of the clinically relevant composite endpoint of cardiovascular death, myocardial infarction and ischemic stroke (MACE), and the individual components of this endpoint. However, the method had a lower sensitivity to detect the more ambiguously registered endpoint of ischemia-driven revascularization. Positive predictive values were low to moderate for all endpoints. Third, we found strong agreement in estimated cumulative incidence rates between the twomethods of collection. Using investigator-reported endpoints or adjudicated endpoints as reference standard did not affect EHR data retrieval accuracy. Fourth, we found that the majority of misidentified endpoints depended on misclassification by the manual operator. Our method of validating routinely collected healthcare data from hospital EHRs from multiple vendors, using data from a clinical outcomes trial as reference standard, has not been described before. The results are similar to previous studies that compared correspondence of routinely collected healthcare data registered in healthcare claim databases to investigator-reported outcomes and endpoint evaluation by an endpoint adjudication committee, in which sensitivity of using routinely collected healthcare data ranged between 49% to 100%. 13,20–23 Our results show sensitivity, specificity, and negative predictive values are particularly high for strictly defined clinical outcomes, such as major adverse cardiovascular events. With these properties, EHR data retrieval from routinely collected healthcare data can be considered to complement or substitute investigator-reported endpoint collection of such outcomes in clinical research. Less strictly defined or ambiguously registered clinical outcomes such as ischemia- driven revascularization show lower levels of accuracy. EHR data retrieval using broadlydefined clinical outcomes is appropriate for clinical research involving large datasets, where estimates are less effected by missed observations. 24 Depending on the context of the clinical research, methods that increase sensitivity could reduce specificity, as was observed in our study with the secondary composite endpoint