503 - A natural language processing (NLP) pipeline for creation of a multi-institutional kidney biopsy registry

Publication Number: 2489.503

Vitaly Lorman, Children's Hospital of Philadelphia, Evanston, IL, United States; Amy Goodwin Davies, Childrens Hospital of Philadelphia, Philadelphia, PA, United States; Rodrigo Azuero-Dajud, Childrens Hospital of Philadelphia, 1822 N Front Street, PA, United States; Mohan Kashyap Pargi, Childrens Hospital of Philadelphia, King of prussia, PA, United States; Laura Finn, Children's Hospital of Philadelphia, Philadelphia, PA, United States; Rebecca Scobell, Childrens Hospital of Philadelphia, Philadelphia, PA, United States; Caitlin J. Alexander, Childrens Hospital of Philadelphia, Philadelphia, PA, United States; Adya Maddox, Childrens Hospital of Philadelphia, Philadelphia, PA, United States; Grace Park, Childrens Hospital of Philadelphia, Philadelphia, PA, United States; Qiwei Shen, Childrens Hospital of Philadelphia, Philadelphia, PA, United States; Aliyah Jones-Griffith, Children's Hospital of Nevada at University Medical Center, Pennsauken, NJ, United States; Peter EF. Camacho, Childrens Hospital of Philadelphia, Philadelphia, PA, United States; Hanieh Razzaghi, Children's Hospital of Philadelphia, Philadelphia, PA, United States; William E.. Smoyer, Nationwide Children's Hospital, Powell, OH, United States; Charles Bailey, Children's Hospital of Philadelphia, Philadelphia, PA, United States; Joseph Flynn, Seattle Children's, Seattle, WA, United States; Caroline Gluck, Nemours Children's Health, Wilmington, DE, United States; Priya S. Verghese, Northwestern University The Feinberg School of Medicine, Chicago, IL, United States; Mark Mitsnefes, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States; Sujan Kumar Mutyala, Texas Children's Hospital, Richmond, TX, United States; Suchitra Rao, University of Colorado School of Medicine, Aurora, CO, United States; Hiroki Morizono, Children's National Health System, Washington, DC, United States; Jonathan S.. Chawla, Baylor College of Medicine, Houston, TX, United States; Michelle Denburg, Children's Hospital of Philadelphia, Philadelphia, PA, United States; H Timothy Bunnell, Nemours Children's Health, Wilmington, DE, United States

Background: PEDSnet is a national, multi-institutional network of 10 children's hospitals and maintains an electronic health record (EHR) database which has been a resource for numerous nephrology studies. The EHR database consists of predominantly structured data such as diagnosis codes and lab results. Histopathological information is typically embedded within free text kidney biopsy reports and unavailable in the structured EHR data.

Objective: We present a natural language processing (NLP) pipeline for the creation of a multi-institutional kidney biopsy registry. This approach enables automated extraction of clinical features from kidney biopsy reports, expanding the existing data to include histopathological information and facilitating future observational studies and trial feasibility assessment and recruitment.

Design/Methods: The NLP pipeline (Figure 1) distinguishes between native and transplant biopsies and extracts 29 renal concepts, including diagnoses of specific glomerulopathies and, in the case of transplant biopsies, whether there is evidence of antibody and cellular-mediated rejection. Renal concept extraction is implemented using a hybrid approach consisting of both pattern matching using terms from a lexicon (developed by nephrologists and pathologists) as well as a large language model for assertion status (e.g. whether a biopsy report concludes a patient is positive or negative for rejection or whether the biopsy report is consistent with a diagnosis). To validate the asserted diagnostic categories, kidney biopsy reports are sampled and reviewed by nephrology/pathology experts.

Results: A pilot round of validation was completed at one site, with the pipeline identifying a correct concept (a correct diagnosis for native biopsies or evidence of rejection for transplant biopsies) in 84% of reports. The pipeline was subsequently expanded to three additional institutions, correctly identifying a correct concept in 90% of reports. Additional performance metrics are reported in Tables 1 and 2.

Conclusion(s): The pipeline has shown promising performance and generalizes well to new institutions. Next steps involve another round of refinement and validation at 5 additional sites. Once finalized, it can be rerun on data refreshes to maintain an up-to-date kidney biopsy registry, anticipated to contain extracted concepts for up to 10,000 kidney biopsy reports. The registry will be a valuable resource for future studies which require histopathologic information from kidney biopsy reports, and the pipeline will be made publicly available as a tool for investigators at other institutions.

Figure 1
NLP pipeline for extraction of renal concepts from kidney biopsy reports

Table 1
Children's Hospital of Philadelphia (CHOP) pilot validation performance (N=257)

Table 2
Validation performance at additional three institutions (N=179)