OBJECTIVE: Fracture risk calculators, such as the Fracture Risk Assessment Tool, or FRAX, typically lie outside the clinician workflow. Conversely, the electronic health record (EHR) is at the center of the clinical workflow, and many variables in the EHR could predict fractures without having to verbally ascertain FRAX risk factors. We sought to evaluate the utility of EHR variables to predict fractures and, as a proof of concept, to create an EHR-based fracture risk model.
METHODS: Routine clinical data from 24,189 subjects presenting to primary care from 2010 to 2018 was utilized. Major osteoporotic fractures (MOFs) were captured by physician diagnosis codes. Data was split into training (n = 18,141) and test sets (n = 6048). We fit Cox regression models for candidate risk factors in the training set, and then created a global model using a backward stepwise approach. We then applied the model to the test set and compared the discrimination and calibration to FRAX.
RESULTS: We found variables related to vital signs, utilization, diagnoses, medications, and laboratory values to be associated with incident MOF. Our final model included 19 variables, including age, BMI, Parkinson\'s disease, chronic kidney disease, and albumin levels. When applied to the test set, we found the discrimination (AUC 0.73 vs. 0.70, p = 0.08) and calibration were comparable to FRAX.
CONCLUSIONS: Routinely collected data in EHR systems can generate adequate fracture predictions without the need to verbally ascertain fracture risk factors. In the future, this could allow for automated fracture prediction at the point of care to improve osteoporosis screening and treatment rates.
目标:骨折风险计算器,如断裂风险评估工具,或者FRAX,通常位于临床医生工作流程之外。相反,电子健康记录(EHR)是临床工作流程的中心,EHR中的许多变量可以预测骨折,而无需口头确定FRAX风险因素。我们试图评估EHR变量预测骨折的实用性,作为概念的证明,建立基于EHR的骨折风险模型。
方法:利用2010年至2018年接受初级保健的24189名受试者的常规临床数据。主要骨质疏松性骨折(MOFs)由医师诊断代码捕获。数据分为训练集(n=18,141)和测试集(n=6048)。我们在训练集中拟合候选风险因素的Cox回归模型,然后使用向后逐步方法创建了一个全局模型。然后,我们将模型应用于测试集,并将辨别和校准与FRAX进行比较。
结果:我们发现了与生命体征相关的变量,利用率,诊断,药物,以及与事故MOF相关的实验室值。我们最终的模型包括19个变量,包括年龄,BMI,帕金森病,慢性肾病,和白蛋白水平。当应用于测试集时,我们发现了歧视(AUC0.73vs.0.70,p=0.08),校准与FRAX相当。
结论:在EHR系统中常规收集的数据可以产生足够的骨折预测,而无需口头确定骨折危险因素。在未来,这可以在护理点进行自动骨折预测,从而提高骨质疏松症筛查率和治疗率.