Real-world comparison between the HealthVCF and HealthOST AI models for the detection of vertebral fractures (VFs) from existing CT scans

Rachel Eckert; Muhammad Javaid

doi:10.1530/obabs.08.OC2.8

OC2.8

Prev Next

Section Contents Cite

Bone Abstracts (2026) 8 OC2.8 | DOI: 10.1530/obabs.08.OC2.8

OP2026 Oral Communications Oral Communications 2: Quality Improvement Projects (15 abstracts)

Real-world comparison between the HealthVCF and HealthOST AI models for the detection of vertebral fractures (VFs) from existing CT scans

Rachel Eckert ¹ & Muhammad Javaid ²

Author affiliations

¹Oxford University Hospitals NHS Foundation Trust, Oxford, United Kingdom;²University of Oxford, Oxford, United Kingdom

Background: Even though VF strongly predict hip fracture risk, most adults with VFs are missed by FLSs. Over 10% of CT scans performed in adults aged ≥50 years have a VF, but often unreported. A number of different AI models have been developed to identify VF from existing CT scans. We aimed to compare the performance of AI-detected VF between the commercially available Nanox-AI HealthVCF (VCF) and HealthOST (OST) models.

Performance characteristics from VCF to OST, including estimated impact per 1000 scans
	AI flag rate	Sensitivity	Specificity	% clinician agreement	NPPV	PPV	Estimated scans	impact using	per 12.7%	1000 rate
								AI flagged scans (n)	AI flagged scans with VF (n)	AI flagged scans without VF (n)	VF missed
HealthVCF	32.7%	88.6%	73.3%	74.7%	98.4%	26.1%	327	112	215	14.5
HealthOST	28.0%	95.7%	81.7%	83.5%	99.2%	43.1%	280	121	159	5.4

Methods: A consecutive series of CT scans, including the thoracic and lumbar spine, from 1st to 3rd October 2025, was analysed with the VCF and OST models using balanced sensitivity-equivalent settings. While VCF analyses axial images to generate a sagittal image output, HealthOST uses the primary sagittal images to measure vertebral height. HealthOST also flags scans with low BMD. An FLS nurse with experience in identifying VFs read the AI-positive and AI-negative CT scans for both VCF and OST, blinded to order, to identify VFs with moderate/severe deformity or a clear endplate fracture. Gwet’s AC was used to test agreement, and Fisher’s exact test was used to test differences between the two AI models.

Results: 364 CT scans were analysed by both models and included in the results. The clinician confirmed VF prevalence was 9.6% with VCF and significantly higher at 12.7% with OST (P 0.001). There was 94.8% agreement (Gwets AC 0.93, P 0.001). 10.1% of OST AI scans were also identified as potentially low BMD with a significantly higher rate in patients with confirmed VFs (23.9% vs 8.2%, P 0.001)

Conclusion: The updated OST model flagged fewer scans for clinical confirmation and missed significantly fewer VFs. The impact of AI-estimated low BMD status on FLS triage and patient management requires further follow-up.