Predicting imminent fractures in patients with a recent fracture or starting oral bisphosphonate therapy: development and international validation of prognostic models

The availability of anti-osteoporosis medications with rapid onset and high potency requires tools to identify patients at high imminent fracture risk (IFR). There are few tools that predict a patient's IFR. We aimed to develop and validate tools for patients with a recent fracture and for patients initiating oral bisphosphonate therapy. Models for two separate cohorts, those with incident fragility fracture (IFx) and with incident oral bisphosphonate prescription (OBP), were developed in primary care records from Spain (SIDIAP database), UK (Clinical Practice Research Datalink GOLD), and Denmark (Danish Health Registries). Separate models were developed for hip, major, and any fracture outcomes. Only variables present in all databases were included in Lasso regression models for the development and logistic regression models for external validation. Discrimination was tested using area under curve (AUC) and calibration was assessed using observed versus predicted risk plots stratified by age, sex, and previous fracture history. The development analyses included 35,526 individuals in the IFx and 41,401 in the OBP cohorts, with 671,094 in IFx and 330,256 in OBP for the validation analyses. Both the IFx and OBP models demonstrated similarly good performance for hip fracture at 1 year (with AUCs of 0.79 [95% CI 0.75 to 0.82] and 0.87 [0.83 to 0.91] in Spain, 0.71 [0.71 to 0.72] and 0.73 [0.72 to 0.74] in the UK, and 0.70 [0.70 to 0.70] and 0.69 [0.68 to 0.70] in Denmark), and lower discrimination for major osteoporotic and any fracture sites. Calibration was good across all three countries. Discrimination and calibration for the 2-year models was similar. The proposed IFR prediction models could be used to identify more precisely patients at high imminent risk of fracture and inform anti-osteoporosis treatment selection. The freely available model parameters permit local validation and implementation.