Developing an OMOP-Standardized Prostate Cancer Database and Improving Data Quality Using NLP and PSA-Based Algorithms

@misc{wang2026developing,
  title = {{Developing an OMOP-Standardized Prostate Cancer Database and Improving Data Quality Using NLP and PSA-Based Algorithms}},
  author = {J. Wang and J. C. Jackson and A. Garza and S. Nalla and T. Ninnemann and Y. Zhang and Y.-F. Kuo},
  howpublished = {medRxiv},
  year = {2026},
  month = jun,
  doi = {10.64898/2026.06.30.26356984},
  url = {https://www.medrxiv.org/content/10.64898/2026.06.30.26356984v2},
  note = {CC BY-NC-ND},
}

TY  - UNPB
TI  - Developing an OMOP-Standardized Prostate Cancer Database and Improving Data Quality Using NLP and PSA-Based Algorithms
AU  - J. Wang
AU  - J. C. Jackson
AU  - A. Garza
AU  - S. Nalla
AU  - T. Ninnemann
AU  - Y. Zhang
AU  - Y.-F. Kuo
PY  - 2026
DA  - 2026/06/30
PB  - medRxiv
DO  - 10.64898/2026.06.30.26356984
UR  - https://www.medrxiv.org/content/10.64898/2026.06.30.26356984v2
AB  - ObjectiveTo develop and evaluate an Observational Medical Outcomes Partnership-(OMOP-) standardized prostate cancer database from the University of Texas Medical Branch (UTMB) Epic Electronic Health Record (EHR) and improve data quality using natural language processing(NLP) and prostate-specific antigen-(PSA) based algorithms.  Materials and MethodsWe built a data pipeline to transform UTMB Epic EHR data for patients identified through the Galveston Cancer Registry (2010-2021) into OMOP Common Data Model (CDM) v5.4. Data quality was assessed by comparing the OMOP-standardized data with Galveston Cancer Registry data using availability agreement, Cohens kappa and Intraclass Correlation Coefficient. NLP was used to extract PSA, Gleason score, and cancer stage from clinical text, and PSA-based algorithms were used to identify missing treatment and biochemical recurrence.  ResultsWe extracted 815 analytic cases from UTMB EHR. Among them, 700 (85.9%) were complete and concordant with the cancer registry. PSA showed excellent value agreement. Structured Gleason score and stage data were sparse (n<20), but NLP greately improved capture. Treatment agreement was good compared with the cancer registry and improved slightly for radical prostatectomy after applying a PSA-based algorithm. Using PSA trajectories, we identified 60 cases of biochemical recurrence.  DiscussionThe OMOP-standardized data from UTMB showed good agreement with the cancer registry. However, structured EHR fields incompletely captured diagnosis, pathology, and treatment details. NLP and PSA-based algorithms substantially improved data capture. Manual review also revealed errors in registry data, showing that OMOP-standardized EHR data can complement and help improve cancer registry quality. Conclusion: OMOP standardization combined with NLP and PSA-based algorithms improved prostate cancer data quality and research readiness.  Lay SummaryElectronic health record (EHR) systems contain valuable long-term health information about cancer, but their use is limited by difficulties in data extraction, the lack of standard data formats, and missing information in structured fields. In this study, we built a data pipeline to convert EHR data into the Observational Medical Outcomes Partnership (OMOP) Common Data Model, a widely used standard format that facilitates data sharing and collaborative research. To address missing data, we used natural language processing (NLP, a computational technique that enables computers to interpret human language) to extract prostate-specific antigen (PSA) levels, Gleason scores (a measure of cancer aggressiveness), and cancer stages from unstructured clinical notes. We also applied algorithms that track PSA blood trends over time to identify unrecorded cancer treatments and detect cancer recurrence. Comparing this database with a cancer registry demonstrated that standardized EHR records provide accurate clinical information and long-term follow-up. While structured EHR fields lacked key tumor characteristics, NLP dramatically improved their capture. Tracking PSA trends identified previously unrecorded treatments and flagged 60 potential cancer recurrences. Overall, combining standardized EHR formats with NLP and PSA-based algorithms creates a rich database to support real-world prostate cancer research.
ER  -