September 2026 â What if your medical record could identify an increased risk of hip fracture years before a fracture ever occurred? A new AI hip fracture prediction tool developed by researchers in Sweden suggests that may be possible.
In a nationwide study involving more than 3.5 million adults aged 50 and older, researchers developed a tool called FRACTURE-ML that uses information already contained in health records to estimate a person’s risk of hip fracture. At two years, the model identified substantially more of the people who went on to experience a hip fracture than a screening approach based on recent fracture history.
That is an impressive finding. But from the Better Bones perspective, identifying risk is only the beginning.
Key takeaways
- FRACTURE-ML, a machine-learning model from the University of Gothenburg, predicts hip fracture risk using only routine medical-record data: diagnoses, prescriptions, procedures, age and socioeconomic information.
- At two years it identified 84% of people who went on to fracture a hip, versus 12% for a recent-fracture screening approach, about seven times as many.
- The trade-off: it also flags more people who will not fracture (79% specificity vs. 98%).
- It has not been validated outside Sweden, lacks lifestyle data such as exercise, smoking and alcohol, and is not a standard screening test in the United States.
What did the researchers do?
Researchers at the University of Gothenburg analyzed data from 3,542,647 Swedish adults age 50 and older. Participants were assigned a baseline date between 2011 and 2013 and followed through the end of 2021. During a median follow-up of just over nine years, 142,327 people experienced a hip fracture.
Instead of requiring a bone density test, office visit, or questionnaire, FRACTURE-ML analyzed routinely collected registry information, including:
- Age and demographic information
- Diagnoses
- Prescription medications
- Medical procedures
- Socioeconomic information
The researchers initially considered nearly 140,000 possible variables. Their full machine-learning model ultimately used 2,500 predictors. Interestingly, a much simpler model using only 35 variables performed nearly as well.
The study was published August 27, 2026, in PLOS Medicine.
How accurate is the FRACTURE-ML hip fracture prediction model?
The full FRACTURE-ML model demonstrated strong ability to distinguish between people who would and would not experience a hip fracture. Its reported area under the curve, or AUC, was:
- 1 year: 0.89
- 2 years: 0.88
- 5 years: 0.85
- 10 years: 0.82
An AUC of 0.5 would be no better than chance, while 1.0 represents perfect discrimination.
The reduced 35-variable model performed almost as well, with an AUC of 0.87 at two years and 0.85 at five years. Traditional statistical models also produced similar results, suggesting that access to large amounts of health information may be at least as important as the use of artificial intelligence itself.
Where does the “seven times more” number come from?
The researchers compared FRACTURE-ML with a Fracture Liaison Service-style approach based on having had a recent fracture. At two years, FRACTURE-ML identified 84% of those who would go on to have a hip fracture, compared with 12% using the recent-fracture approach. That is approximately seven times as many people identified.
There is an important trade-off, however. FRACTURE-ML had a specificity of 79%, compared with 98% for the recent-fracture approach. In other words, identifying many more people who will fracture also means flagging more people who will not.
That distinction matters enormously when deciding what should happen after someone is labeled “high risk.”
Why this research is exciting
One of the biggest challenges in osteoporosis is that bone loss can progress silently. People often do not know they have compromised bone health until a bone density test reveals it â or, worse, until a fracture occurs.
A system that can identify increased risk from information already in a person’s medical record could prompt earlier evaluation and prevention.
The researchers suggest that people identified as being at elevated risk could then receive further assessment, including evaluation of physical function and balance and a DXA bone density test. Being found at high risk of fracture would also encourage a careful evaluation of all medications being used that might increase fracture risk.
But this is where an important question arises: Once we identify someone as being at risk, what do we do with that information?
What this model cannot see
For all the information contained in Sweden’s national registries, several important pieces of the bone-health picture were missing. The researchers specifically acknowledge that they did not have lifestyle information such as:
- Exercise
- Smoking
- Alcohol intake
And a medical database alone cannot give us a complete picture of a person’s nutrition, strength, balance, fall risk, digestive health, nutrient status, stress, or many of the other factors that can influence bone health. This is especially important because many of these factors are modifiable.
A risk-prediction model can tell us that a problem may be developing. It cannot necessarily tell us why that individual is losing bone or what combination of factors is contributing to that risk.
Important limitations
This is promising research, but FRACTURE-ML is not yet a routine clinical screening tool.
- It has not been externally validated. The model was developed and tested using Swedish registry data. The researchers state that validation in other countries and healthcare systems is still needed.
- Implementation has not been tested. We do not yet know what happens when a healthcare system uses this type of screening in real-world practice.
- Lifestyle information was missing. Smoking, exercise, and alcohol intake were not available in the national registry data.
- Machine learning was not dramatically better than conventional statistics. Traditional Cox regression models produced similar results, suggesting that much of the model’s power may come from having access to a very large and detailed dataset.
The authors themselves conclude that external validation and implementation studies are necessary before the clinical usefulness of FRACTURE-ML can be established.
Dr. Brown’s take: Prediction should be the beginning, not the end
I welcome any tool that helps us recognize fracture risk before the fracture happens.
For decades, I have encouraged women to become proactive about their bone health rather than waiting for osteoporosis to progress or for a fracture to tell us that something is wrong. If technology can help us identify people who deserve a closer look earlier, that could be very useful.
But identifying risk is not the same thing as understanding the cause of that risk. This study illustrates that distinction beautifully.
What a medical record cannot tell us
The computer can analyze diagnoses, prescriptions, procedures, age, and thousands of other pieces of information contained in a medical record. Yet some of the questions I would immediately want to ask a person are not necessarily captured there:
What are you eating? Are you getting enough of the key nutrients needed to build and maintain bone? Are you exercising in ways that stimulate bone and maintain muscle? How is your balance? Have you been falling? Are there digestive issues interfering with nutrient absorption? Are there medical, hormonal or metabolic factors contributing to excessive bone loss? Is your body chemistry pH in optimal balance? Do you live with an excessive dietary acid load?
Risk identification should start a workup, not a prescription
Those questions matter because my approach has always been to look beyond the diagnosis of osteoporosis and ask why this particular person is losing bone. This is also why I would not want a high-risk prediction to become an automatic pathway from computer algorithm to prescription.
Medication can be appropriate for some individuals, particularly when fracture risk is very high, and that decision should be made thoughtfully with a healthcare professional. But risk identification should also be an invitation to conduct a thorough medical workup and investigate the many potentially modifiable contributors to bone health.
Why an earlier warning matters
There is another lesson here that I find encouraging. The model does not know everything about you. It does not know what you are going to eat tomorrow. It does not know whether you are going to begin strength training, improve your balance, correct a nutrient insufficiency, address an underlying medical problem, stop smoking, or make other changes that support your skeleton.
A risk estimate is not destiny.
So, I see research like this as potentially giving us an earlier warning system. And earlier warning gives us something extremely valuable: time to investigate, time to intervene, and time to build stronger bones. That is where I believe the real opportunity lies.
The Better Bones bottom line
FRACTURE-ML represents an intriguing shift in osteoporosis screening: instead of waiting for someone to enter the healthcare system for a bone density test or after a fracture, existing medical information could potentially help identify people who need attention sooner.
But prediction alone does not prevent fractures.
The next step should be understanding the individual’s complete bone-health picture â including bone density when appropriate, fracture history, medical causes of bone loss, bone turnover, nutrition, exercise, balance and fall risk, and other modifiable factors.
Technology may become increasingly sophisticated at telling us who is at risk. The most important question will still be: What can we do about it?
Frequently asked questions about AI hip fracture prediction
What is FRACTURE-ML?
FRACTURE-ML is a machine-learning model developed at the University of Gothenburg in Sweden that estimates a person’s risk of hip fracture using information already in their health records, such as diagnoses, prescriptions, procedures, age and socioeconomic data. It does not require a bone density scan or questionnaire.
Can my medical records predict a hip fracture?
In a study of more than 3.5 million Swedish adults aged 50 and older, routine medical-record data predicted hip fracture with an AUC of 0.88 at two years. The tool has not yet been validated outside Sweden, so it is not known how well it would work in other healthcare systems.
How is it different from FRAX or a bone density test?
FRAX and DXA bone density testing require a clinical visit, questionnaire answers or a scan. FRACTURE-ML works automatically from existing records, so it could flag people for a closer look before they ever see a clinician about their bones. It is meant to prompt further assessment, not replace it.
Is FRACTURE-ML available in the United States?
No. FRACTURE-ML is a research tool and is not currently a standard clinical screening test in the United States.
Primary source
Axelsson KF, Litsne H, Konstantinou K, Khalid H, Pivodic A, Lorentzon M. “A clinical decision support tool for accurate hip fracture prediction: A nationwide cohort study.” PLOS Medicine. Published August 27, 2026. DOI: 10.1371/journal.pmed.1005190.
Important: FRACTURE-ML is a research tool and is not currently a standard clinical screening test in the United States. Individual fracture risk and osteoporosis treatment decisions should be discussed with a qualified healthcare professional. Do not start, stop, or change a prescription medication based on a risk score or news report.
