Finding Silent Fractures: The Economics and Evidence of AI Spine Screening

Sam: Imagine an AI tool flags a hidden spinal fracture with a specificity above 90%, but the economic model assumes only 15% of the people identified actually end up getting treatment. Where is the disconnect?

Maya: That is exactly the puzzle at the heart of a new document published by the National Institute for Health and Care Excellence, or NICE. It is called 'Artificial intelligence (AI) technologies to aid opportunistic detection of vertebral fragility fractures: early value assessment', published on 14 October 2025.

Sam: Before we dive in, we should tell everyone that our voices are AI-generated. This podcast is a summary of a public document, and we are here to unpack what it actually means for clinical practice and for health technology companies.

Maya: Right. And this document, which replaces the older HTE34 guidance, is looking at a very specific, very common problem: vertebral fragility fractures, or VFFs. These happen when the bones in the spine become weaker than normal, usually because of osteoporosis.

Sam: The document notes these fractures can happen from low-energy trauma, like just a fall from standing height, or even spontaneously from normal day-to-day activities. It is not always a massive injury.

Maya: Exactly. And the patient experts in the document were very clear that these fractures can be life changing and emotionally challenging. They cause deformity, height loss, immobility, and pain. They also increase the risk of death and are a strong predictor for future osteoporotic fractures, like hip fractures.

Sam: But a lot of them are missed, right? The clinical experts mentioned that despite efforts to raise awareness, most VFFs remain undiagnosed. That is where the AI comes in.

Maya: Yes. Millions of diagnostic images are taken in the NHS every year for completely unrelated reasons. The idea is 'opportunistic detection'. The AI looks at X-rays or CT scans that include the spine but were ordered for something else, and flags if there is a fracture.

Sam: So NICE evaluated several AI tools for this. And they essentially split them into groups. 5 get a tentative green light for NHS use, and 3 are sent back to the research bench. Let us start with the 5 that made the cut.

Maya: The 5 technologies recommended for use in the NHS during a 3 years evidence generation period are BriefCase-Triage from Aidoc Medical, CINA-VCF Quantix from Avicenna.AI, HealthVCF and HealthOST from Nanox AI, and IB Lab FLAMINGO from IB Lab.

Sam: And the key distinction here is that all 5 of those are designed to detect fractures on CT scans. Chest, abdominal, or pelvic CTs. But the 3 that didn't make the cut are focused on X-rays. Why did NICE draw the line there?

Maya: It comes down to the clinical workflow and how X-rays are used in standard care. Let us look at BoneView by Gleamer and TechCare Spine by Milvue. They analyse spinal X-rays. But the clinical experts explained that if someone is getting a spinal X-ray, it is usually to investigate back pain.

Sam: Meaning the radiologist is already paying very close attention to the spine.

Maya: Exactly. Because the spine gets a thorough review in those cases, it is much less likely a fracture would be missed. So the committee felt it is uncertain if the AI offers additional benefit there. Plus, the document notes there is no clinical evidence at all for TechCare Spine.

Sam: Okay, but what about the other one? Annalise Enterprise CXR, and its counterpart Annalise Container CXR. That one analyses chest X-rays. Surely a chest X-ray isn't focused on the spine, so AI could help there?

Maya: That is a fair point, and it can use anterior-posterior or posterior-anterior views. But the clinical evidence submitted for the Annalise technology was heavily based on studies that mostly or only included lateral chest X-rays. Those are side-view images.

Sam: And are those common in the NHS?

Maya: No. The clinical experts noted that lateral chest X-ray images are no longer commonly done in the NHS. The added cost and radiation exposure usually outweighs the benefit. The company said some trusts still do them for specific high-risk groups, but the committee concluded the evidence might not be generalisable to normal NHS practice.

Sam: So for now, Annalise Enterprise CXR, BoneView, and TechCare Spine can only be used in research. Access has to be through company, research, or non-core NHS funding. That is a massive takeaway for companies building in this space. If your training data relies on an imaging view the health system rarely uses, your AI might not get funded.

Maya: Precisely. The clinical setting matters just as much as the algorithm's performance. And speaking of performance, the external assessment group, or EAG, looked at 22 studies for the clinical-effectiveness review.

Sam: Let us break down those 22 studies. Where was the bulk of the evidence coming from?

Maya: Most evaluated HealthVCF with 8 studies, Annalise with 5 studies, IB Lab FLAMINGO with 4, and CINA-VCF Quantix with 3. BoneView and BriefCase-Triage had 1 study each. And during consultation, 3 more were submitted, including 2 on HealthOST.

Sam: And the results? You mentioned earlier they had a specificity above 90%.

Maya: Yes, mostly for detecting moderate and severe vertebral fractures when compared against a reference standard. But here is the catch. The committee judged that evidence comparing these tools directly against standard care in the NHS is very limited and uncertain.

Sam: Because standard care in the NHS would be a human radiologist reporting on the scan, right?

Maya: More specifically, a specialist committee member explained the true reference standard in the NHS would be at least 1 radiologist with specialist musculoskeletal training reviewing the diagnostic image specifically looking for a VFF.

Sam: That makes sense. But opportunistic screening means the scan is being read by a reporting practitioner who isn't necessarily a musculoskeletal specialist, and who is probably looking for lung or abdominal issues. So standard care might actually miss quite a lot.

Maya: Exactly. 13 studies did compare the AI against standard care, mostly the original radiology report, but most were not done in the UK. Still, the committee agreed the technologies are likely to improve detection rates. They are just not sure by exactly how much.

Sam: I noticed the document mentions that 4 companies—Aidoc Medical, Annalise.AI, Nanox AI, and Avicenna.AI—have settings to control the AI software's sensitivity and specificity at setup or during use. That seems useful for hospitals wanting to manage their false positives.

Maya: It is. False positives are a massive concern here. The clinical experts warned that false positives could lead to unnecessary dual-energy X-ray absorptiometry, or DEXA, scans. And if a lot of extra images need review by a radiologist, that creates a significant workforce impact.

Sam: Right, the downstream resource impact. Identifying more fractures sounds great clinically, but operationally, it means more specialist reviews and more DEXA scans. The document actually notes there are already capacity issues across the NHS for both acquisition and reporting of DEXA scans.

Maya: And that is exactly why this is an early value assessment requiring 3 years of evidence generation. NICE wants companies to track the impact of identifying additional VFFs on referral rates for other services, on treatment, and on the workload of healthcare professionals.

Sam: Let us talk about the economics, because that really dictates whether this is viable. The EAG built an early economic model with a 1-year time horizon. How much are these AI tools actually costing per scan?

Maya: The company-provided costs were commercial in confidence, so the EAG used a hypothetical scenario with a generic AI technology costing £7.36 per scan. The committee noted that the cost per scan of all technologies was similar to or below that £7.36 generic cost.

Sam: Except for BoneView, right? No cost was provided for that one.

Maya: Correct. For BoneView, the EAG added a notional additional cost of £1 per scan. The committee pointed out it is unknown if that reflects the true cost, making BoneView's potential cost effectiveness even more uncertain.

Sam: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Maya: And now, back to the document.

Sam: So we have the cost of the software. But what about the clinical benefit in the model? If you find a fracture, you treat it, and the patient's quality of life improves. How did they measure that?

Maya: There was a quality-of-life benefit modelled, but it was based on a study of people who were diagnosed after presenting with symptoms. The EAG felt that people diagnosed opportunistically might not have symptoms that are as severe.

Sam: Ah, so they assumed the benefit of treating an opportunistic fracture would be lower.

Maya: Yes, to be safe, the EAG actually halved the utility value in the base case to better reflect this population. The patient experts pushed back a bit, noting that VFFs can be extremely painful and symptoms are sometimes mistaken for something else, like stomach pain.

Sam: That is fascinating. The patients are saying, 'just because you found it by accident does not mean I am not in pain'. But the model is conservative. And what about that 15% treatment figure I mentioned at the start?

Maya: That was a major area of uncertainty. The model assumed that only 15% of people correctly identified as having a VFF would have treatment. This was based on a small expert elicitation study. The committee wants to know how many people are already on osteoporosis treatment when the VFF is found, because for them, finding the fracture doesn't provide any added benefit in reducing future risk.

Sam: And the model only had a 1-year time horizon. That seems really short for a bone disease.

Maya: The committee agreed it was a major limitation. Many benefits, like preventing future hip fractures, happen beyond one year. But the EAG said longer-term modelling was too uncertain due to the lack of evidence on clinical management changes. They do expect that including longer-term costs and benefits would likely improve the cost effectiveness of the technologies.

Sam: Even with those limitations—halving the quality of life gain, only assuming a 15% treatment rate, and ignoring long-term fracture prevention—the committee still concluded it is plausible the AI technologies could be cost effective.

Maya: Yes, in the base case they were more expensive than standard care, but they led to quality-of-life gains for the people identified and treated.

Sam: Let us shift to the real-world operational risks of buying and using these tools. The document talks a lot about 'failure rates'. What does that actually mean in this context?

Maya: 11 of the studies reported failure rates, meaning the AI software failed to process a diagnostic image or failed to produce a definitive report. The reasons included 'unable to upload image', 'unable to process image', 'incorrect classification as unsupported anatomical regions', 'unsuccessful model inference', and 'unable to analyse image'.

Sam: That is a lot of different ways for the software to crash or refuse to read the scan. A specialist committee member mentioned it could just be related to poor image quality. But it means hospitals need to factor in that not every scan sent to the AI will get a result.

Maya: Exactly. And for companies, the committee explicitly listed failure rates and the reasons for failure as a gap where further evidence should be generated.

Sam: I also want to touch on the software lifecycles. There was an interesting detail about Nanox AI having both HealthVCF and HealthOST on the list. Why have both?

Maya: HealthOST is essentially the updated version of HealthVCF, trained on a similar but larger dataset. At the time of the meetings, both were available in the NHS, but the regulatory approval for HealthVCF is due to expire in 2028. It is unlikely to be on the UK market after that.

Sam: So Nanox is gradually replacing it. The committee noted their preference is for evidence to be generated using HealthOST, since it is the one that will stick around. That is a great reminder for hospital buyers to check the expiration date on a product's CE mark before signing a multi-year contract.

Maya: Definitely. And speaking of regulatory approvals, the guidance mandates that these technologies can only be used as long as they have appropriate regulatory approval, including NHS England's Digital Technology Assessment Criteria, or DTAC, approval.

Sam: Let us look at who these tools are actually for. We know osteoporosis is common in older people, particularly women, trans men, and non-binary people after menopause. Are there age restrictions on the AI itself?

Maya: Yes, there are. 4 of the tools—CINA-VCF Quantix, HealthVCF, HealthOST, and IB Lab FLAMINGO—are indicated as Class IIa or IIb devices specifically for people 50 years and over. BriefCase-Triage, on the other hand, is Class IIa for people 18 years and over.

Sam: And Annalise Enterprise CXR is Class IIb for people 16 years and over, while BoneView is Class IIa for people over 2 years. But if the disease mostly affects people over 50, what happens if you run the AI on a younger person?

Maya: That is a huge system risk. If you run it on everyone, the resource impact from false positives could be massive. The prevalence of VFFs is lower in people under 50, so screening them is less beneficial. However, younger people can have VFFs due to other risk factors.

Sam: Right, the document lists secondary causes like rheumatoid arthritis, inflammatory bowel disease, malabsorption, chronic corticosteroid or glucocorticoid use, low body mass index, smoking, and alcohol intake. A younger person with inflammatory bowel disease on long-term steroids might absolutely have a spinal fracture.

Maya: Exactly. But the problem is that AI technologies can only be configured to assess specific diagnostic images based on demographic information, like age, not on those specific clinical risk factors. You cannot easily tell the AI to only screen the younger people who take steroids.

Sam: That is a huge limitation of current hospital IT. The AI does not read the patient's full medical history; it just gets the DICOM file from the picture archiving and communications system, the PACS. So hospitals have a choice: screen all younger people and drown in false positives, or only screen over 50s and miss the high-risk younger patients.

Maya: Which is why the committee concluded that evidence should still be generated across all age and risk groups to see if screening younger people represents value for money. They also highlighted a risk around equality and AI training data.

Sam: Bias in the algorithm.

Maya: Exactly. There is a lack of transparency about the data used to train the algorithms. The mean or median ages in the clinical evidence studies were generally between 65 and 80 years. The committee warned the AI might perform worse in populations underrepresented in the training set, like ethnic minorities or younger people with comorbidities.

Sam: So companies need to be transparent about their training data, and hospitals need to watch for reduced diagnostic accuracy in diverse populations.

Maya: Yes. And another clinical risk the committee addressed was the fear of deskilling. If radiologists rely on the AI, will they lose the ability to spot these fractures themselves?

Sam: What did the clinical experts say?

Maya: They argued there is no imminent risk of deskilling. In fact, they thought healthcare professionals might actually learn from the AI's feedback. Plus, these tools are strictly decision aids; they do not replace the definitive radiology review.

Sam: Alright, let us recap the big picture. 5 AI tools that read CT scans have been conditionally recommended by NICE for a 3-year period to help find hidden vertebral fractures. 3 tools that read X-rays need more research.

Maya: For clinicians, this means you might soon see more incidental fractures flagged in your PACS viewers, but there is no clear pathway everywhere in the NHS for non-acute cases, meaning fracture liaison services might get overwhelmed.

Sam: And for companies, the lesson is clear: your AI might have over 90% specificity, but if you cannot prove how it changes standard NHS care, or if it causes too many false positives that bottleneck DEXA scanners, it will be hard to get permanent funding.

Maya: A reminder to our listeners that this podcast is a summary of a public document and is not medical advice. We have linked the full NICE guidance in the show notes so you can explore the recommendations and the evidence generation plan yourself.

Sam: Thanks for joining us. We will see you next time on Smart Summaries.