(SLA-059) Data stacking of RWD from two EMR Databases in the United States using Tokenization

Background: Stacking real-world data (RWD) databases comprised of different patient populations can increase the size of patient cohorts and representativeness of the data. This can be particularly beneficial where patient numbers are low in a desired study cohort or where the exposures or outcomes are rare. The steps in this process involve (1) an assessment of database eligibility for stacking and tokenization/harmonization (2) tokenization to identify and remove duplicate patients from one source and (3) subsequent harmonization into a common data model (CDM).

Objectives: To assess the degree to which the overall size of a treated metastatic breast cancer (mBC) patient cohort increases through use of tokenization of two distinct treated mBC patient cohorts.

Methods: This study used RWD from two nationwide electronic health record based deidentified databases, ConcertAI and Flatiron Health. Tokenization (i.e., the process of replacing sensitive data with non-sensitive placeholders called “tokens”) was done using proprietary software from Datavant to assign an encrypted token to personally identifiable information (PII). Tokenization was done independently by Flatiron and ConcertAI using tokens generated by the Datavant tokenization software. No PII was shared with Datavant or between the databases. Encrypted tokens were applied to combinations of the following variables: first name, last name, administrative gender, and date of birth (DOB). Distinct counts of two tokens designs were used to identify unique patients across the two databases. For the duplicate patients identified, only one record was retained for the stacked data.

Results: In total there were 38,975 mBC patients in the Flatiron Health Database and 39,373 mBC patients from ConcertAI’s data lake. All mBC patients in Flatiron were tokenized with 1,603 mBC patients with specific treatment criteria applied in ConcertAI for overlap assessment. Of the tokenized Flatiron patients, 1247 met specific treatment requirements and there was an overlap between the two databases of 248 patients. The final sample included patients with curated data from ConcertAI (N=593) and all treatment-specific patients from Flatiron Health Database (N=1,247), including overlapping patients (N=248) resulting in an approximate sample size increase of 50%.

Conclusions: The absolute increase in patient number in the combined data demonstrates the usefulness of this method where patient numbers are low. Tokenization is a key step to enable deduplication when stacking datasets, increasing sample size while protecting patient privacy.