Summary
Real-world data (RWD) provides valuable insights into breast cancer (BC) treatment patterns and outcomes. Multiple RWD sources collect information on the same medical data, thus combining them results in a more diverse population. Tokenization enables removal of duplicate patients, so datasets can then be harmonized into a single cohort. However, traditional schema harmonization is complex and time-consuming. Objective: To apply a scalable, AI-driven approach to harmonize ConcertAI (CAI) and Flatiron Health Research Database (FHRD) BC cohorts into a common data schema, enabling the creation of a unified, expanded patient cohort.
Methods
The Cornerstone AI platform (CSAI) —purpose-built for RWD transformation and cleaning—was used to convert CAI BC data into the FHRD schema by leveraging its large language models to map CAI tables and fields into the FHRD target schema (post tokenization by Datavant software). Differences in variable names were standardized using custom embedding models and, where necessary, data from multiple CAI tables were transformed into derived tables to match FHRD structure. An interactive UI allowed adjustments tailored to downstream analytic needs.
Results
Harmonization enabled a 45% increase in final sample size (596 CAI; 1,339 FHRD; n=1,935) after deduplication of the 8.7% duplicate patients across CAI and FHRD. Of 36 CAI tables, 28 (78%) tables were harmonized to 15 FHRD tables (the remainer represented CAI tables with no FHRD analogue, e.g. imaging, adverse events, early-stage disease, sites of metastasis). This included successful mapping of 278 CAI fields into 153 FHRD fields. Three FHRD harmonized tables required intermediate derived tables. e.g. the FHRD schemas cancer diagnosis and staging tables contained information from three CAI tables, where conditional logic selected appropriate fields and records. Line of therapy and socioeconomic status were too divergent to harmonize.
Conclusions
AI-enabled harmonization of BC RWD across CAI and FHRD is feasible. It increases patient cohort size therefore statistical power, and supports downstream analytical code reuse, all while maintaining transparency —key requirements for clinical research and healthcare applications.
