Oral Presentation 29th Australian and New Zealand Bone and Mineral Society Annual Scientific Meeting 2019

Natural Language Processing of radiology reports for the identification of patients with fracture (#59)

Nithin Kolanu 1 2 3 , Shane Brown 4 , Amanda Beech 3 5 , Jackie Center 1 2 6 , Christopher P White 1 3 5 6
  1. Garvan Institute of Medical Research, Sydney, NSW
  2. Endocrinology, St Vincent's Hospital, Sydney, NSW
  3. Endocrinology/ Diabetes, Prince of Wales Hospital, Randwick, NSW, Australia
  4. Abbott Diagnostics, Macquaire Park
  5. Royal Hospital for Women, Randwick, NSW
  6. University of New South Wales, Sydney, NSW, Australia

Fracture liaison services address the treatment gap for those with osteoporosis (OP) who fracture and are not treated. Screening high volumes of radiology reports for fractures with Natural Language Processing (NLP) could identify patients that have not been recognized or treated. This study is an analytical and clinical validation of XRAIT (X-Ray Artificial Intelligence Tool) at its development site (Prince of Wales Hospital, Sydney) and external validation in an adjudicated cohort from the Dubbo Osteoporosis Epidemiology Study (DOES).

Methods: XRAIT performs a Boolean search of radiology reports for fracture and fracture related terms. It can be trained for site specific reporting styles. At the development site, XRAIT and an independent, blinded fracture liaison clinician (manual review) were used to review the emergency patient presentations of people over 50 years of age during the same 3-month period. XRAIT analyzed plain radiographs and CT scans (n = 5089) while manual review of clinical records was possible for n =228. External validation: XRAIT was used to analyze digitally readable radiology reports in an untrained cohort from DOES (n = 371) to calculate sensitivity and specificity.

Results: XRAIT identified a 6-fold higher number of potential significant fractures (433/5089) compared to manual case finding (72/228). 418 were true positives (96.5%) and 257 (61.5%) were acute fractures. Only 29% were started or recommended anti-resorptive therapy and these included those seen by fracture liaison service. XRAIT unadjusted for the local radiology reporting styles in DOES had a sensitivity of 69.6% and specificity of 92.3%.

Conclusion: XRAIT identifies clinically significant fractures. Its high specificity even in an untrained cohort suggests it could be used at other sites. Automated methods of patient identification may assist fracture liaison services to identify fractures that still remain largely untreated.

  1. Grundmeier, R. W., A. J. Masino, T. C. Casper, J. M. Dean, J. Bell, R. Enriquez, S. Deakyne, J. M. Chamberlain, and E. R. Alpern. "Identification of Long Bone Fractures in Radiology Reports Using Natural Language Processing to Support Healthcare Quality Improvement." [In eng]. Appl Clin Inform 7, no. 4 (Nov 9 2016): 1051-68. https://doi.org/10.4338/aci-2016-08-ra-0129.
  2. Do, B. H., A. S. Wu, J. Maley, and S. Biswal. "Automatic Retrieval of Bone Fracture Knowledge Using Natural Language Processing." [In eng]. J Digit Imaging 26, no. 4 (Aug 2013): 709-13. https://doi.org/10.1007/s10278-012-9531-1.