Systems based identification of patients with vertebral fractures using natural language processing — ASN Events
  1. Schedule
  2. Poster Viewing I (odd numbers)
  3. Systems based identification of patients with vertebral fractures using natural language processing

Systems based identification of patients with vertebral fractures using natural language processing (#203)

Jay Pandya 1 , Kirtan Ganda 2 , Markus J Seibel 2
  1. Bone Research Program, ANZAC Research Institute, Sydney, NSW, Australia
  2. University of Sydney, Concord, NSW, Australia

Fracture prevention programs exist worldwide to address the ‘care gap’ in osteoporosis management. These programs primarily identify patients with symptomatic non-vertebral fractures. However, asymptomatic vertebral fractures are not systematically identified, despite conferring a higher risk of re-fracture. This study aims to develop a Natural Language Processing (NLP) method to systematically identify patients with radiographically verified vertebral fractures via searching free-text electronic radiology reports and to determine its clinical utility. The study consisted of two phases:

  1. The Development Phase used twelve search terms to identify patient reports with vertebral fractures. Each report was reviewed to confirm the presence or absence of vertebral fractures. The total output (number of reports extracted from the search term), and positive predictive value (reports verified as fracture relative to output) was then calculated.
  2. The Implementation Phase, applied the three most effective search terms from phase one to identify patients with confirmed vertebral fractures. The patients were then invited to attend the SFP program for further management as a part of best practice.

The Development phase revealed three search terms with the highest total output: ‘Loss of Height’, ‘Compression Fracture’ and ‘Crush Fracture’. During the Implementation Phase, 126 reports were identified, representing 96 individual patients. 69 patients (72%) had a vertebral fracture and were referred to a SFP program. The term ‘Loss of Height’ was more effective in identifying patients with vertebral fractures compared to ‘Compression Fracture’ and ‘Crush Fracture’. The term ‘Compression Fracture’ was similar to the term ‘Crush Fracture’ in its ability to identify vertebral fractures.

In conclusion, simple NLP methods can be utilised to identify patients with vertebral fractures via electronic radiology report searches. These NLP methods may be translated to other SFP programs to identify patients with vertebral fractures, thereby further narrowing the ‘care gap’ in osteoporosis management.