Leveraging Electronic Health Records for Phenotyping - Translational Informatics: Realizing the Promise of Knowledge-Driven Healthcare - page 60

successes in genetic studies and clinical informatics arose from secondary data use.
Researchers in the discovery of BRCA1 used the Utah Population Database, which
includes familial histories linked to data extracted from electronic health databases
[ 11 ]. Evans et al. demonstrated one of the fi rst examples of a learning health system
with the Antibiotic Assistant, which used data in the EHR from similar patients to
compute recommendations for antibiotic prescribing [ 19 ]. These studies were suc-
cessful in part because they used very specifi c data for defi ned purposes among
defi ned populations. As more data were collected and the data types, use and scope
broadened, researchers began to face limitations in secondary use of EHR data.
Early research using EHR data for improving adverse drug events (ADEs) showed
that structured data were incomplete and underestimated the number of ADEs[ 20 ].
Studies in data mining were also limited by methods extracting data from EHRs
[ 21 ]. Issues of data quality and completeness have continued to challenge EHR data
reuse [ 22 , 23 ]. For many years, the early potential was unmet as researchers identi-
fi ed and faced multiple barriers, even while the collection of data was increasing [ 6 ].

4.3 Recent Developments

Now that we have established how the use of EHR data for phenotypes has emerged
over time, and why it is interesting now, we can discuss how it is done more clearly.

4.3.1 Extracting Data from EHRs

EHRs contain various types of data that are used for various purposes. Some data
are collected primarily for administrative purposes. Demographics information is
needed to identify an individual, both for treatment and payment. Diagnosis data
indicate the overall condition of a patient, and are generally collected for billing.
Procedure data indicate the various actions taken by clinicians, and are also used for
billing. Because demographics and billing data are most commonly and consis-
tently collected for patients, they have been a primary source for phenotype infor-
mation in population research [ 24 ]. However, because they are used primarily for
non-clinical care, researchers have observed inconsistencies and errors in billing
data [ 2 ]. Other data are collected primarily to support clinical care. These include
medications prescribed, assessments made, tests and activities ordered, results of
tests, actions performed, and statements of clinical judgment. Laboratory test results
and standard patient assessments (e.g., vital signs) are generally stored in structured
form, and can be used to interpret phenotypes for diseases that are specifi cally indi-
cated by test results [ 25 ]. Medication orders and prescriptions are also often struc-
tured, but may also be stored as part of unstructured text (e.g., medication history).
Medications can also reveal patient clinical conditions by what is being treated [ 26 ].
The richest source of clinical information is usually stored as unstructured text in