Natural Language Processing to Improve Vertebral Fracture Identification in Older Adults Receiving Osteoporosis Therapy.
Authors
Affiliations (10)
Affiliations (10)
- Division of Research, Kaiser Permanente Northern California, Pleasanton, CA, USA.
- Department of Health Systems Science, Kaiser Permanente Bernard J Tyson School of Medicine, Pasadena, CA, USA.
- Department of Spine Surgery, The Permanente Medical Group, Kaiser Permanente Oakland, Oakland, CA, USA.
- Department of Medicine, University of Washington School of Medicine, Seattle, WA, USA.
- Geriatrics Research Education & Clinical Center, Minneapolis VA Healthcare System, Minneapolis, MN, USA.
- Department of Medicine, University of Minnesota, Minneapolis, MN, USA.
- J Harold Harrison, MD, Distinguished University Chair in Rheumatology, Division of Rheumatology, Department of Medicine, Medical College of Georgia at Augusta University, Augusta, GA, USA.
- Edward Hines, Jr Veterans Administration Hospital, Hines, IL, USA.
- Endocrinology and Metabolism Section, Richmond Veterans Affairs Medical Center, Richmond, VA, USA.
- Division of Endocrinology, Diabetes and Metabolism, Virginia Commonwealth University, Richmond, VA, USA.
Abstract
Vertebral fracture identification using administrative codes is challenging because diagnoses often occur outside hospital and orthopedic care settings. Outpatient diagnoses improve capture but increase misclassification. Accurate methods are important for population-level research and care. To develop an accurate natural language processing (NLP) algorithm to apply to radiology reports for vertebral fractures. Adult members of an integrated health care delivery system who were aged 50-89 years at bisphosphonate initiation (1998-2019) and members 3 years later (index) were followed through 2023 for an inpatient, emergency, or outpatient visit with a vertebral fracture diagnosis code (N = 19,444). There were 14,172 (72.9%) with a radiology report from spine imaging ≤ 90 days before or after the vertebral fracture diagnosis encounter. Using a random subset of 219 diagnosed potential cases with radiology reports that were physician-adjudicated for vertebral fractures, the authors developed and validated a rule-based NLP algorithm for confirming vertebral fracture that was applied to qualifying radiology reports for the population. Compared with physician adjudication based on radiology reports, the final NLP algorithm had a positive predictive value of 95.7% and a negative predictive value of 100% for vertebral fracture. When the NLP algorithm was applied to radiology reports from 14,172 adults with a vertebral fracture diagnosis, 88.5% were classified as vertebral fracture. This proportion was similar (90.2%) for the subset of 10,917 adults without prior vertebral fracture diagnosis. These findings demonstrate that NLP applied to radiology reports provides a scalable and accurate method to improve vertebral fracture identification for population-based research and surveillance using electronic health record data.