1034 — Electronic Clinical and Administrative Data Can be Used to Predict Patients at Risk of Fragility Fracture

LaFleur J (Salt Lake City VAMC), Nelson R (Salt Lake City VAMC), Pickard SG (Salt Lake City VAMC), Nebeker JR (Salt Lake City VAMC)

Objectives:
Prospective epidemiology studies have shown that patient-reported risk factors for fracture can predict fractures with acceptable discrimination. One model based on prospectively-collected clinical risk factors alone was shown to have a C-statistic of 0.71 in women; bone mineral density (BMD) alone had a C-statistic of 0.79. Our objective was to determine whether risk factors from VA clinical and administrative datasets could predict fragility fractures with similar discrimination.

Methods:
We used Veterans Health Administration clinical and administrative datasets for veterans treated in the US Rocky Mountain region to create a cohort of males over the age of 49 who received care during the calendar years of 2005 and 2006. We identified constructs related to fracture risk factors in the clinical and administrative datasets including age, sex, race/ethnicity, body mass index (BMI), smoking, alcohol use, as well as several comorbid diseases and drug exposures. We used Cox Proportional Hazards regression to construct a prognostic model for identifying patients at highest risk of fracture. The model was developed using a stepwise procedure that considered both clinical knowledge about causal associations as well as the significance of statistical associations.

Results:
A total of 85,191 patients were included in the final analysis cohort. The mean age of the cohort was 66.9 years. Constructs related to risk factors that were also strong predictors of fracture in multivariable models were prior fragility fracture (0.8%), smoking (19.8%), alcohol use disorder (2.1%), BMI < 18.5 (1.0%), and glucocorticoid use (1.7%). The incidence of fragility fracture (non-traumatic, non-pathologic fractures of the hip, wrist, or spine) over a mean follow-up time of 33 months was 2.0 fractures per 1,000 person-years, occurring in 450 veterans. The final model had a C-statistic of 0.7913 and a bootstrap shrunken C-statistic of 0.7908.

Implications:
A model created using constructs related to risk factor for fracture predicts patients at risk of fracture with acceptable discrimination. C-statistics were similar to those found for BMD and for a risk-stratification rule created using prospectively collected risk factors.

Impacts:
This work has important implications for creating targeted clinical reminders for identifying patients at highest risk of fracture.