FRACTURE-ML flags hip-fracture risk from registry data

Before a clinician measures body mass index or asks about lifestyle, a new model could estimate a patient's hip-fracture risk from existing records. Developed by researchers Kristian Axelsson and Mattias Lorentzon at the University of Gothenburg in Sweden, FRACTURE-ML identified nearly seven times more people at risk within two years than current Swedish clinical screening, without an in-person assessment.

The model was built from Swedish national registry data covering 3,542,647 people aged 50 and older, followed for up to 10 years. During that period, 142,327 participants sustained a hip fracture. FRACTURE-ML used more than 100,000 variables from diagnoses, medications, procedures, demographic information and socioeconomic data to distinguish people who later fractured a hip from those who did not.

Its performance held up in a separate group not used to develop the model. The area under the curve — a measure of how well a prediction separates higher-risk from lower-risk people — was 0.89 one year ahead and 0.85 five years ahead. Against existing Swedish screening, sensitivity rose from 0.12 to 0.84, while specificity fell from 0.98 to 0.79. A simpler version using 35 variables performed nearly as well, suggesting that a more practical tool may retain much of the larger model's accuracy.

So what changes, concretely? Health services could use routinely collected data to find more people who may benefit from preventive measures before a fracture occurs, rather than relying only on assessments that require direct contact. That could make population-level screening more efficient for older adults, whose hip fractures are associated with disability, illness and death.

The limits are clear. Registry data do not capture lifestyle factors such as smoking and alcohol use, which may affect fracture risk. The results come from Sweden, and the authors say FRACTURE-ML must still be validated in other countries and tested in real-world care. They also note that carefully developed traditional statistical models achieved similar accuracy, making the advance less about one algorithm than about using comprehensive health data more effectively.