AI predicts hip fracture risk based on EHR

Clinical takeaway: Hip fracture risk may soon be predictable before the first fracture, without adding clinic workload.  

Identifying who will break a hip is one of the more consequential unsolved problems in preventive care. Hip fractures carry high rates of disability, loss of independence, and death in older adults, and effective preventive measures exist, from osteoporosis medications to fall prevention. But those measures only work when they reach the right people. 

Risk calculators such as FRAX and QFracture depend on information a patient supplies in person, including body mass index, smoking, and alcohol use. That requirement has kept population-wide screening out of reach in practice. The prevailing approach instead waits for a first fracture and works to prevent the second. A Swedish research team examined whether the record trail patients already leave behind could provide an early clue, and built one of the largest studies ever conducted on the question to find out. 

Working from national registry data on Swedish adults aged 50 and older, the team's deep-learning model, FRACTURE-ML, built on 2,500 variables proved highly accurate at separating people who went on to fracture a hip from those who didn't. One year ahead, it correctly ranked a future fracture patient as higher risk than a non-fracture patient 89% of the time, and that discrimination held up well at five years, at 85%. Its predicted probabilities also matched what actually happened: people the model scored as high risk fractured at close to the predicted rates. 

A stripped-down version using just 35 variables, led by age, sex, prescription count, and marital status, performed nearly as well, and traditional statistical models were almost equivalent to the machine-learning approach. The predictive power was based on the breadth of the data, not the algorithm itself. 

Swedish screening activates only after a first fracture, and by that standard it catches about 12 of every 100 people headed for a hip fracture within two years. The model caught 83 of 100, nearly seven times as many, at the cost of flagging more people who would not fracture: it cleared 78% of non-fracture patients versus 98% under current practice. The authors present the two approaches as complementary rather than competing, with the model adding a primary prevention layer that current practice lacks. 

The study drew on Sweden's linked national registers, which capture hospital diagnoses, prescriptions, procedures, and demographic and socioeconomic data for every resident. All 3.5 million adults were 50 or older, had no recent osteoporosis medication, and were followed for up to ten years, during which 142,327 sustained a hip fracture. The team started with nearly 140,000 candidate variables, winnowed them to the 2,500 most predictive for the full model, and validated performance in a holdout group of more than 350,000 people kept separate from model development. 

What stands between FRACTURE-ML and clinical use is specific: the model has never been tested outside Sweden, and no study has yet examined what happens when its predictions are acted on. The authors sketch how implementation could look, with a national service running risk scores from existing registry data and flagging elevated-risk individuals for bone density measurement, physical function assessment, and treatment where warranted.  

The harder question they leave to health systems is where to set the threshold, since flagging more future fractures means evaluating far more people. That tradeoff, they suggest, could shift with age, favoring specificity in younger patients and sensitivity in older ones. 

"The findings show that it is possible to predict hip fracture risk at the population level without direct patient interaction. This approach could help target preventive measures more efficiently and potentially reduce the number of hip fractures," said Kristian Axelsson, MD, PhD, researcher at the Sahlgrenska Osteoporosis Centre, University of Gothenburg, Sweden.

Source: Axelsson KF, et al. (2026 Aug 27) PLoS Med. A clinical decision support tool for accurate hip fracture prediction: A nationwide cohort study