An enhanced Fracture Risk Evaluation Model (FREM) using national health data on morbidity and medications

Osteoporosis is a major health concern in older individuals. Efficient case-finding is essential for timely risk assessment and treatment for high-risk patients. To prevent fractures and reduce the risk of subsequent disability, approaches offering clinically sufficient sensitivity with acceptable specificity are warranted. Although pharmaceutical osteoporosis treatment is effective, it is often diagnosed at a late stage, for instance, following a fracture. The aim of this study was to extend the existing Fracture Risk Evaluation Model (FREM), which identifies individuals at risk of an imminent (1-yr) major osteoporotic fracture (MOF) based on administrative health data. This extension (FREMVer2) included data on morbidity and medications and evaluated age-specific risk cut-offs to enhance the risk assessment of MOF and hip fracture (HF) risk, respectively. We included the entire population of Denmark aged ≥45 yr at baseline (2022; N = 2 493 180), not previously diagnosed with osteoporosis or receiving osteoporosis treatment. The cohort was divided into 4 groups stratified by sex and age (<65 yr, ≥65 yr). Each of the 4 groups was randomly split into a 60% development, a 20% model validation, and a 20% cut-off validation cohort. All diagnoses from Danish hospitals and filled prescriptions from Danish pharmacies from 2007 to 2021 were used as possible predictors for MOF. These predictors correspond to information that in Denmark is automatically transferred to general practitioner's electronic health records; hence, prediction would be possible in general practice. Models were constructed by logistic regression with LASSO regularization, determining the preferred regularization hyper parameter by cross-validation and forcing categorical age to be included. Across subgroups, the models obtained poor to acceptable area under the curves (AUCs) of 0.656-0.714 for MOF, and acceptable AUCs of 0.728-0.764 for HFs. Additionally, the models achieved sensitivities of around 80% or higher in almost all subgroups. This performance, together with the available predictors, makes FREMver2 a feasible decision support system as a step toward an opportunistic screening program in health care settings with access to administrative data.