SEATTLE — A machine learning-based clinical decision-support tool predicted hip fracture risk from electronic medical records in Sweden with a significantly higher sensitivity than fracture liaison service (FLS).

The tool has a significant advantage over other approaches in predicting fracture risk because it does not require patient contact, said Mattias Lorentzon, MD, PhD, chief physician at the Osteoporosis Clinic at Sahlgrenska University Hospital, Gothenburg, Sweden, during presentation of the results at the American Society for Bone and Mineral Research (ASBMR) 2025 Annual Meeting.

The machine-learning model predicted 2-year fracture risk with an area under the curve (AUC) of 0.88 compared with 0.55 for the commonly used FLS approach of screening people who have had a fracture in the past year.

“This high accuracy was largely sustained over longer follow-up periods and after accounting for competing risk of death. And finally, the implementation of [machine learning] has potential to be used for general population screening or implemented in electronic health records, which would allow automated risk prediction that would enhance the identification of high-risk patients. And by doing this, we could target more of the high-risk patients and reduce hip fracture rates,” said Lorentzon.

Can It Be Applied to Patients in the Physician’s Office?

Claus Glüer, PhD, recently retired professor of medical physics at Kiel University in Kiel, Germany, who attended the presentation, called the study “very, very important” during the Q&A session. He endorsed the value of the approach for population studies but questioned Lorentzon as to whether the predictive factors would be applicable to individual patients in the physician’s office. Lorentzon acknowledged the concern and said that one of the next steps for the research is to test the tool against bone mineral density and other commonly used individual risk factors to “see if it accurately captures those who would benefit from treatment.”

Another attendee asked Lorentzon about the need to validate the results in other populations. “We are hoping in the near future to be able to run our algorithm and test it in the Danish registered data set, which is very similar to the Swedish, so we hope to have external validation soon,” he replied.

But even the best population databases may be missing key data such as family history of hip fractures, Glüer said in an interview. He noted that it’s unclear if the Swedish database includes such information, although it “seems to be one of the most complete ones.”

He also wondered further about the applicability of the results to the physician’s office. “The strongest risk factors that we typically encounter in osteoporosis, like prior fractures, certain diseases like Parkinson’s, et cetera. Probably you capture that person’s risk by just looking at the one, two, or three strongest risk factors in that individual — much simpler and much more individualized. I’m not sure whether all the candidate predictors that are in this population-based model would be the same or the most relevant ones if you’re dealing with an individual patient. It remains to be tested,” Glüer said.

He endorsed the approach for use in policy decisions. “You get the overall picture in the entire population, including prevalence of the disease and prospects to really calculate the burden of the disease. And also, maybe you can estimate how much you could do if you change your policy by providing earlier treatment,” he added.

Study Results

In the study, the researchers analyzed data from all individuals living in Sweden between 2011 and 2013 who were at least 50 years old and had not previously been exposed to osteoporosis medicine (n = 3,542,647). All participants were followed up until the end of 2021.

They analyzed national register data that included fractures, comorbidities, surgery, socioeconomic factors, mortality, and prescriptions, including a total of 140,131 variables that encompassed diagnoses, medications, procedures, and demographics. The researchers divided the population into discovery (25%), development (65%), and validation (10%) cohorts. There were 142,327 hip fractures over an average of 9.04 years of follow-up.

The machine-learning model, using 2500 variables, predicted 1-year fracture risk with an AUC of 0.89, a 2-year AUC of 0.88, a 5-year AUC of 0.83, and a 10-year value of 0.82. A simplified model with 35 variables performed nearly as well, with corresponding AUC values of 0.88, 0.87, 0.85, and 0.81.

The most influential variables for prediction were age, sex, number of prescriptions, marital status, alcohol-related disorders, anti-dementia drugs, and hospital stay days in the previous 2 years. FLS-based screening for fracture risk yielded a 2-year AUC of 0.55.

At 2 years, the machine-learning method would have identified about seven times more individuals at risk for fracture than the FLS approach (sensitivity, 0.84; 95% CI, 0.82-0.85 vs 0.12; 95% CI, 0.11-0.13), although with a lower specificity (0.79; 95% CI, 0.79-0.79 vs 0.98; 95% CI, 0.98-0.98).

The study was funded by the Swedish Research Council, the Sahlgrenska University Hospital Research Funds, and the Gothenburg Medical Society. Lorentzon declared receiving lecture or consulting fees from Amgen, UCB, Janssen, Astellas, Alexion, Consilient Health, Parexel International, Medac, Medison, Pharmacosmos, and Gedeon Richter. Glüer declared having no relevant financial disclosures.

Jim Kling is a writer based in Bellingham, Washington.