Automated Identification of Vertebral Compression Fractures Reported by Radiologists on Routine Chest Radiography
Supervisor / Principle Investigator:
MD Class of 2019
Introduction: Vertebral compression fractures (VCF) are a hallmark sign of osteoporosis. however, their often-incidental presence on chest x-ray rarely initiates treatment. Large scale manual search for vertebral fractures is time-consuming and impractical. We tested the feasibility of an automated search for VCFs found in radiologist-generated reports.
Methods: The last 30,000 emergency room and outpatient reports between 3 hospitals were searched. Patients aged 60+ with reports including postero-anterior and lateral films were included. Reports were automatically flagged if they contained keywords suggesting VCF. A musculoskeletal fellowship-trained radiologist then performed a blinded manual review of 100 randomly selected flagged radiographs and 100 randomly selected non-flagged radiographs. The program search results were compared to those of the manual review.
Results: 2213 unique cases (7.4%) were flagged for VCF, with the search term “compression fracture” being the most commonly used identifier. The 200 randomly selected cases yielded a mean age of 76.7 years, with 84 males (43%) and 114 females (57%). Manual review of these cases by a radiologist yielded an 90.0% auto-capture accuracy rate (k=0.80, 95% confidence interval 0.72-0.88). Of the 20 manually reviewed discrepant cases, the auto-capture program provided 8 false-negative and 12 false-positive results. The program had 92% sensitivity and 89% specificity.
Conclusions: Our search program had a 92% sensitivity and 89% specificity for detecting VCFs, yielding 7.4% patients positive for VCF. 92% of patients flagged by the automated search did indeed have VCFs and treatment could be considered in these patients. This program forms the basis for potential further prospective inquiry into osteoporosis treatment initiation.