P0014 - Deriving a De-Identified Dataset from EPIC-Derived Data for Use in Clinical Research: Pearls and Pitfalls in Database Creation and Analysis

Background: Given the ubiquity and amount of information available through EHRs, there is treasure trove of data available for research purposes if we can gather and analyze this information in a systematic way. In 2018, EPIC announced plans for Cosmos, a large scale collaborative research database pooling EHR data from multiple institutions likely to be widely adopted for research purposes1. Pediatrics is a field where it often difficult to obtain high power studies and enroll patients for prospective trials. Through our institutional database we compared the efficacy of lactate, CRP, procalcitonin in predicting mortality among adults in an effort to validate our database.

Methods: The study sample was obtained from an EPIC-derived database created from all patients seen at Loma Linda University Medical Center between January 2013 and March 2019. De-identified unique hospital encounters with at least 1 serum or arterial/venous blood gas lactate drawn in patients 18 or older were used to limit the study subset of critically ill patients. Subsequently age, gender, hospital length of stay, admission vital signs, admission labs, C-reactive protein (CRP) and procalcitonin when available for these hospital encounters.

Results: The study population included 58,172 distinct hospital encounters based on study inclusion criteria. Binomial logistic regression models were run individually for Lactate, CRP and Procalcitonin due to differences in number of patients available for each measure. Male gender, Lactate, INR were significantly associated with mortality with odds ratios of 1.253, 1.293, 1.283 in the lactate logistic regression model. CRP and Procalcitonin were significantly associated with mortality in their respective models but with clinically insignificant odds ratios at nearly 1 with 1.028 and 1.008 respectively. Receiver Operator Curves showed areas under the curve of 0.670, 0.553, 0.672 and inflection points of 1.7, 8.0 and 0.3 for lactate, CRP, and procalcitonin respectively.

Conclusion: This study offers proof that data from EPIC can be extracted and the data and is suitable for clinical research. We found significant associations between mortality and lactate which has been documented in numerous studies with a ROC curve inflection point of 1.7 correlating with values in previous studies reporting normal values ranging from under 1.4 to 2.3 2. Implementation of an EHR derived database is feasible but does require an initial investment of time and resources. After the initial learning curve however, the payoff is an infrastructure for further research. Any pediatrician in the institution can now perform high power studies and answer a variety of clinical questions that otherwise be challenging to investigate. There are limitations in using such data notably a lack of standardization of when labs were drawn for a particular admission since such data is not protocolized unlike conventional prospective clinical research.