AI catches overlooked broken vertebra on X-rays, with sizable cost-saving implications

An artificial intelligence model can accurately catch overlooked broken vertebra on chest X-rays, potentially saving patients and the healthcare system money in the process. 

Such vertebral compression fractures are a common manifestation of osteoporosis, with chest radiography an opportunity to spot this concern, initiate treatment and prevent further fractures. Only about 21% of women 50 to 64 are screened for osteoporosis despite an 8.4% prevalence in this age group, experts noted. 

Scientists with Massachusetts General Hospital in Boston tested the use of a chest X-ray triage software from Annalise.ai, sharing their findings Wednesday in the Journal of the American College of Radiology [1]. AI accurately identified vertebral compression fractures with a sensitivity of 89.3% and specificity of 89.2%. 

“The performance of this model achieved the benchmark sensitivity and specificity of 80% that are often used by the FDA for computer-assisted triage devices,” Bernardo C. Bizzo, MD, PhD, a radiologist and senior director of Mass General Brigham’s AI business, and co-authors wrote Sept. 17. “These results were mostly maintained across most sex, age, race, ethnicity, and manufacturer subgroups suggesting robust and generalizable performance of the model.”

For the study, Bizzo and colleagues used the AI triage software to retrospectively review a consecutive collection of nearly 600 chest X-rays, which included both frontal and lateral projections. The images came from four U.S. hospitals and were gathered between 2015 and 2021. Up to three thoracic radiologists assessed each radiograph for the presence of vertebral compression fractures, and the AI model then performed inference on the cases. Researchers also reviewed patient charts for the presence of osteoporosis-related ICD-10 codes and prescribed medications one year following the study period. 

The AI model successfully completed its assessment in 595 cases (or 99.8%). Of those, 272 (45.7%) were positive for fractures while 323 came up negative (54.3%). Out of the 236 true-positive cases correctly identified by AI, only 86 (or 36.4%) had a diagnosis of a vertebral compression fracture documented in patient records. About 140 (59.3%) had a diagnosis of either osteoporosis or osteopenia, and only 78 (33.1%) were receiving a disease-modifying medication for this concern. 

Bizzo and co-authors highlighted significant cost implications stemming from their study. Osteoporosis drug Alendronate costs as little as $0.80 for a weekly 70 mg table, while hip fracture treatment ran about $50,508 in 2014. One modeling study estimated that secondary-fracture prevention services could save about $418 per patient. 

“A key pragmatic consideration for the current model is that it could potentially opportunistically screen patients at a smaller cost than other interventions given it automatically operates ‘in the background’ on chest radiographs performed for any/other reasons,” Bizzo et al. wrote. “Care with a bone health or fracture liaison service program could subsequently be automatically initiated. Indeed, the current study suggests a likely substantial opportunity amongst the true-positive patients (i.e., cases where the model correctly identified a vertebral compression fracture) for osteoporosis-related diagnostic coding and medication use where appropriate.”

“Vertebral fractures without spinal cord injury” also is a hierarchical condition category as part of the Centers for Medicare & Medicaid Services risk-adjustment model, the authors noted.  

“This category means that its inclusion as a patient diagnosis through ICD-10 coding can impact a patient’s expected healthcare costs,” Bizzo et al. wrote. “In certain payment models, such as value-based care or capitation, the non-inclusion can result in a medical practice receiving a smaller payment.”