Swedish ML model predicts hip fracture risk 7x better than current screening

Swedish researchers have developed a machine-learning system that can identify people at high risk of hip fracture using only the medical records already stored in national health databases—no doctor's visit, no measurements, no lifestyle questionnaire required. The tool, called FRACTURE-ML, outperforms the screening methods currently used in Swedish hospitals by a striking margin: it catches seven times as many at-risk people within a two-year window.

The study, published in PLOS Medicine on August 27, drew on a decade of health data from 3.5 million Swedish adults over age 50. During the follow-up period, 142,327 of them broke a hip. Kristian Axelsson and Mattias Lorentzon of the University of Gothenburg, working with colleagues, fed the system more than 100,000 variables extracted from diagnoses, medications, procedures, and demographic and socioeconomic information. The algorithm learned to recognize patterns in that vast dataset that predicted who would fracture and who would not.

When tested on people the model had never seen before, FRACTURE-ML achieved an area under the curve—a standard measure of predictive accuracy—of 0.89 when forecasting fractures one year ahead, and 0.85 when looking five years out. Those are strong numbers. But the practical advantage lies in sensitivity: the model identified 84 percent of people who would go on to fracture, compared to just 12 percent caught by current clinical screening. The tradeoff is specificity—the model flags more false alarms, correctly identifying 79 percent of people who will not fracture, versus 98 percent for existing methods. In other words, FRACTURE-ML casts a wider net, and some people it flags will never break a hip. But it misses far fewer people who actually will.

What makes this approach feasible at scale is that it requires no in-person assessment. Hip fractures carry enormous weight in older age: they bring disability, hospitalization, loss of independence, and often death. Yet current risk prediction tools demand that patients come in for measurements and interviews—a barrier that makes population-wide screening impractical. FRACTURE-ML sidesteps that entirely. It works from data that health systems already collect and store.

The researchers also tested a stripped-down version using only 35 variables instead of 100,000. It performed nearly as well as the full model. That finding matters because it suggests the core predictive power might be achievable with something simpler and more transparent than a deep-learning black box—perhaps even a traditional statistical model. The authors note that the advance may be less about the algorithm itself and more about making better use of comprehensive, routinely collected data.

The limitations are real. The model has no information on smoking, alcohol use, or other lifestyle factors that affect fracture risk, because those details don't live in the registry. The system was built and tested on Swedish data; whether it works equally well in other countries, with different health systems and populations, remains unknown. And no one has yet run a real-world trial to see whether hospitals can actually use this tool to prevent fractures—whether identifying high-risk people and intervening before injury actually reduces the number of breaks.

Axelsson and Lorentzon frame the potential clearly: a tool that works from existing data could shift hip-fracture care from reaction to prevention. Instead of treating people after they break, clinicians could target preventive measures—physical therapy, medication, home modifications—to those most likely to fall and fracture. The next step is validation in other countries and implementation studies that test whether the prediction translates into fewer broken hips.