Hip fractures in older people often arrive as a sudden turning point, the moment a life that seemed stable tips toward dependence and loss. In a Swedish study of more than three and a half million adults, researchers have asked whether that turning point can be seen coming without ever sitting a patient down for measurements. Their tool, built from national registry records, anticipates who is likely to break a hip with striking accuracy, raising a quiet question about how medicine might shift from waiting for the fall to watching the path toward it.
Registry-Only Machine Learning Model Predicts Hip Fractures With 0.89 AUC in Nationwide Swedish Study
Hip fractures in older adults carry substantial morbidity and mortality; the study aims to prevent them through earlier identification of at-risk individuals.
Only 24,533 of 100,000 people would be flagged for further evaluation
Mark
What's new here? Hip fracture risk tools already exist.
Mimi
Most of them need measurements like height, weight, smoking, alcohol. This one uses only existing national registry data, so it can score everyone over 50 without anyone being examined.
Mark
And how well does it work?
Mimi
A 1-year AUC of 0.89 and 0.85 at five years, in a holdout cohort of the 3.5 million adults studied. A 35-variable version reached 0.88.
Luke
Worth noting that QFracture and CFracture reported similar or higher numbers, 0.89 and 0.91 in women. The claim is highest among models without direct patient measurements, not highest overall.
Mimi
Fair. The authors make that narrower claim themselves.
Mark
How does it compare with Fracture Liaison Services?
Mimi
It flagged about seven times more at-risk patients within two years and over twelve times more within ten, with a small loss in specificity. The authors call the approaches complementary.
Luke
That comparison is within the study's own data. It shows identification, not fractures actually prevented.
Mark
What about the screening burden?
Mimi
At 80% sensitivity in 100,000 people, 24,533 would be flagged and 460 of 2,300 fractures missed. At that setting, 43 to 64% of women would need intervention.
Luke
And nobody has run this in a real health system yet. Whether flagged people accept scans or treatment, and whether fractures fall, is unknown.
Le Pouls
- A hip fracture can end independence overnight, and this study asks whether that danger can be spotted years earlier.
- The new tool needs no patient examination, sweeping past the traditional need for weight, height and lifestyle measurements.
- Researchers are weighing risk thresholds across Sweden's 21 healthcare regions, where sensitivity and specificity trade against limited resources.
- The tool is promising on paper, but its real-world screening performance remains untested and awaits clinical trials.
Hip fractures in older people often arrive as a sudden turning point, the moment a life that seemed stable tips toward dependence and loss. In a Swedish study of more than three and a half million adults, researchers have asked whether that turning point can be seen coming without ever sitting a patient down for measurements. Their tool, built from national registry records, anticipates who is likely to break a hip with striking accuracy, raising a quiet question about how medicine might shift from waiting for the fall to watching the path toward it.
A team at the University of Gothenburg and Sahlgrenska Academy has built a hip fracture prediction tool that never requires examining the patient. Led by Kristian F. Axelsson, with Mattias Lorentzon among the senior authors, the study was published in PLoS Medicine. It draws on national Swedish registry data covering 3,542,647 adults aged 50 and older.
Called FRACTURE-ML, the tool began by testing 139,980 candidate variables before validating its models in a holdout cohort. The strongest version reached a 1-year AUC of 0.89 and held at 0.85 over five years, which the authors describe as the most accurate registry-based hip fracture model to date. A streamlined 35-variable version performed nearly as well, at 0.88, making practical use more realistic.
Compared with Fracture Liaison Services, which focus on preventing repeat fractures after a first break, the new tool identified far more at-risk patients, with only a marginal loss in specificity. The authors see the two as complementary, with the new approach aimed at primary prevention.
The hardest question is where to set the risk threshold. Authors suggest stricter specificity for younger people and greater sensitivity for older ones, and they acknowledge that each region may choose differently based on resources. Their vision is a national service that calculates scores, which people could view through an app, an online form or their clinician. Those flagged would be offered balance assessments and bone density scans. Still, the study establishes prediction accuracy in registry data; how it performs in actual screening has yet to be tested.
A research team at the University of Gothenburg and Sahlgrenska Academy has built a hip fracture prediction tool that never needs to examine a patient. The study, led by Kristian F. Axelsson with Mattias Lorentzon among the senior authors, was published on August 27, 2026 in PLoS Medicine. It draws entirely on national Swedish registry data covering 3,542,647 adults aged 50 and older.
The tool is called FRACTURE-ML. The researchers started with an unconditional approach, testing 139,980 candidate variables, and then developed and validated models in a holdout cohort that had been kept untouched. The best-performing version used a deep-learning method called DeepSurv with 2,500 variables. Its 1-year AUC was 0.89, and it held at 0.85 over five years. The authors describe it as the most accurate registry-based hip fracture model to date. A reduced version using only 35 variables reached a 1-year AUC of 0.88, which the authors say supports practical use. Traditional Cox models performed almost as well in calibration and discrimination.
The authors place the result against earlier tools. A Danish model, FREM, built on hospital diagnoses from 2,495,339 people, reached AUCs of 0.87 in women and 0.85 in men for one-year hip fracture prediction. They note it covered only the short term, used no machine learning and did not account for competing risk of death. A deep-learning study on electronic health records from over a million people reached 0.81, but its black-box design limited interpretability. QFracture and CFracture, using English and Welsh general practice data, scored 0.89 and 0.91 in women and 0.86 and 0.89 in men. FRAX without bone density scored 0.85 and 0.82. Those tools depend on measured weight, height and lifestyle information, which requires patient contact and is not available in Swedish national data.
The team compared FRACTURE-ML with Fracture Liaison Services, the coordinator-based programs that Swedish authorities give top priority for preventing repeat fractures. FRACTURE-ML identified roughly seven times more at-risk patients within two years and more than twelve times more within ten years, with only a marginal loss in specificity. The authors frame the two as complementary: Fracture Liaison Services act after a first fracture, while the new tool would serve primary prevention.
The central practical question is where to set the risk threshold. In a simulation, an 80% sensitivity setting would require intervention in 43 to 64% of women, depending on age group. An 80% specificity setting would involve only 16 to 17%, at the cost of missing more fractures. The authors suggest higher specificity for younger ages and higher sensitivity for older ones. They note that choices may differ across Sweden's 21 autonomous healthcare regions, depending on resources and willingness to fund follow-up.
They also work through an example. Among 100,000 people over five years, a 2.3% incidence implies about 2,300 hip fractures. At 80% sensitivity, 1,840 would be correctly identified and 460 missed. Only 24,533 people would be flagged for further evaluation, about a 75.5% reduction compared with evaluating everyone.
In the authors' vision, a national service linked to the registries would calculate scores. With consent, people could see them through a health-service app, an online form or a clinician. Those at elevated risk would be offered assessment of physical function and balance and hip and lumbar spine bone density scans, leading to fall prevention or osteoporosis treatment. Swedish registers identify hip fractures with a positive predictive value of 95%.
The study was funded by the Swedish Research Council and several hospital and medical society grants, and the funders had no role in the work. Lorentzon reports consulting or lecture fees from several pharmaceutical companies, outside this work. Code is public on GitHub, though the underlying data are restricted under Swedish confidentiality law. What the paper establishes is prediction accuracy in registry data; how the tool performs when actually deployed in screening has not yet been tested.
Citations marquantes
The approaches should be seen as complementary rather than competitive, with FRACTURE-ML screening added to existing Fracture Liaison Services.— Axelsson and colleagues, PLoS Medicine
Envie de l'histoire complète ?
Lire l'original sur PLOS (Public Library of Science) ↗
Health & Medicine