PREDICTING FUTURE VERTEBRAL COMPRESSION FRACTURES: TESTING THE HYPOTHESIS THAT CHEST RADIOGRAPHS ARE BETTER THAN DEXA

Title: PREDICTING FUTURE VERTEBRAL COMPRESSION FRACTURES: TESTING THE HYPOTHESIS THAT CHEST RADIOGRAPHS ARE BETTER THAN DEXA

Status: Accepted for Presentation

Activity: Scientific Presentations

Control Number: 17280

Control/Tracking Number: 2025-SP-17280-RSNA

Date/Time of Session: 11/30/2025 12:15:00 PM-11/30/2025 12:45:00 PM

Author Block: Anushree M. Burade, MBBS, Parisa Kaviani, MD, Elahe Hosseini, JavierContreras Yametti, MD, Lina Karout, MD, Minjee Kim, Jeong Min Song, MD, Subba R.Digumarthy, MD, Mannudeep K.S. Kalra, MD

Abstract:

Purpose: Extremely low rate of DEXA based bone mineral density (BMD) estimation leads to substantial under-diagnosis of osteoporosis/osteopenia, which are key risk factors for predicting vertebral compression fractures (VCFs). We tested the hypothesis that a three-class classification chest radiograph (CXR)-based AI tool can outperform DEXA-based prediction of future VCFs based on 15-site data.

Methods and Materials: Our IRB-approved, multi-center, external dataset outcome prediction study included 934 chest radiographs from 934 patients (mean age: 71.9 ±11.9 years; female: male 745:189) belonging to one of the 15 US sites (primary care,community care, and quaternary care sites). The study inclusion criteria were age> 21years, posterior-anterior projection CXRs performed within 6 months of DEXA scan, DEXAreports with T and Z scores, and follow-up chest CT or CXRs 3 months to 10 years after the CXR/DEXA exams. We excluded pediatric patients, those with prior spinal surgeries,moderate to severe scoliosis and pacemaker/AICD, traumatic fractures. We used anNLP/CLU-based radiology report data warehouse and search engine to identify consecutive eligible patients. All CXRs were processed with the Promedius AI tool (SouthKorea), not trained with any US data, to obtain probability scores for normal BMD,osteopenia, and osteoporosis. Next, we classified patients into these three-classes based on T and Z-scores from DEXA reports. Data were analyzed to obtain ROC AUCs and oddratios (OR) for predicting VCFs on follow-up chest CT and/or CXRs.

Results: Among the 934 patients, 241/934 (25.8%) developed one or more VCFs on follow-up imaging. The OR for predicting VCFs with CXR-AI (OR 3.15; 95% CI: 1.95-5.08was significantly higher than for DEXA (OR 1.68; 95% CI: 1.13-2.50) (p<0.001). Likewise,the ROC AUC for predicting VCFs with CXR-AI (AUC 0.643, 95% CI: 0.603 - 0.682) was also higher than for DEXA (p<0.01).

Conclusions: Opportunistic screening of BMD with the assessed CXR-based AI model has the potential for predicting future vertebral compression fractures.

Clinical Relevance/Application: Given the underuse of DEXA relative to CXRs, CXR-AIopportunistic screening can represent a viable, initial

Attachment File

Acceptance Letter ([Abstract Acceptance Letter] MGH_VCF.pdf)

Abstract Summary([Abstract Summary] MGH_VCF.pdf)