University of Gothenburg researchers built a machine-learning tool that spots people at risk of a hip fracture years before it happens, drawing only on data already stored in their national health records. It identified roughly seven times more at-risk individuals than the clinical screening in use today.
The model, called FRACTURE-ML, pulls diagnoses, prescriptions, and demographic details already on file. No bone density scan, no questionnaire, no clinic visit. It was published August 27 in PLOS Medicine.
For a two-year prediction window, FRACTURE-ML had a sensitivity of 0.84, compared with 0.12 for Sweden's current fracture liaison services. The team trained the model on 3,542,647 people aged 50 and older, followed for a median of nine years, during which 142,327 hip fractures were recorded.
What the model actually reads
The full model drew on about 2,500 predictors pulled from Sweden's national registries, including hospital diagnosis codes, drugs from the Prescribed Drug Register, and socioeconomic data from Statistics Sweden. A slimmed-down 35-variable version performed almost as well. Accuracy held up from one year out to five, with an area under the receiver-operating curve of 0.89 at one year and 0.85 at five.
Also read: What your gut bacteria could start eating when you don't give them enough fiber
"FRACTURE-ML accurately identified people at high risk of hip fracture using routinely collected health care and population data," said Mattias Lorentzon of the University of Gothenburg in a statement about the study.
His colleague Kristian Axelsson said the findings show it is possible to predict hip-fracture risk at the population level without direct patient interaction.
Why the numbers matter
Hip fractures are among the most consequential falls in later life. The comparison group in the Swedish study, the country's fracture liaison service, is designed to identify high-risk patients for prevention but reached only 12 percent of those who went on to break a hip within two years. FRACTURE-ML reached 84 percent, at the cost of some specificity, meaning more people would be flagged who never fracture.
Mortality after a hip fracture remains steep. A recent US analysis of national death records found the age-adjusted hip-fracture mortality rate fell from 37.07 per 100,000 in 1999 to 23.91 in 2023, a 35.5 percent drop over two decades, though the study still counted 334,905 hip-fracture-related deaths among adults 65 and older during that period.
Also read: How much should you be able to overhead press? The standards by age and sex
FRACTURE-ML has not been rolled out for patient use. The authors say the next step is to test whether feeding the model's high-risk list back into clinical care can actually cut fracture rates.
Also read: Is breakfast really the most important meal? Here's what the evidence actually says
This article is made and published by Jesper Bengtson, who may have used AI in the preparation.