Prediction of Osteoporotic Fractures in Elderly Individuals: a Derivation and Internal Validation Study Using Healthcare Administrative Data

In Canada and other countries, osteoporosis is monitored as part of chronic disease population surveillance programs. While fractures are the principal manifestation of osteoporosis, very few algorithms are available to identify individuals at high risk of osteoporotic fractures in current surveillance systems. The objective of this study was to derive and validate predictive models to accurately identify individuals at high risk of osteoporotic fracture using information available in healthcare administrative data. More than 270,000 men and women aged 66 and older were randomly selected from the Quebec Integrated Chronic Disease Surveillance System. Selected individuals were followed between fiscal years 2006-2007 and 2015-2016. Models were constructed for prediction of hip/femur and major osteoporotic fractures for follow-up periods of 5 and 10 years. A total of 62 potential predictors measurable in healthcare administrative databases were identified. Predictor selection was performed using a manual backward algorithm. The predictive performance of the final models was assessed using measures of discrimination, calibration and overall performance. Between 20 and 25 predictors were retained in the final prediction models (e.g. age, sex, social deprivation index, most of the major and minor risk factors for osteoporosis, diabetes, Parkinson's disease, cognitive impairment, anemia, and anxio-depressive disorders). Discrimination of the final models was higher for the prediction of hip/femur fracture than major osteoporotic fracture and for prediction for a 5- than 10-year period (hip/femur fracture for 5 years: c-index = 0.77, major osteoporotic fracture for 5 years: c-index = 0.71, hip/femur fracture for 10 years: c-index = 0.73, major osteoporotic fracture for 10 years: c-index = 0.68). The predicted probabilities globally agreed with the observed probabilities. In conclusion, the derived models had adequate predictive performance in internal validation. As a final step, these models should be validated in an external cohort and used to develop indicators for surveillance of osteoporosis. This article is protected by copyright. All rights reserved.