SAT-527 - Prediction of Hip Fracture in Diabetic Patients Based on Electronic Health Records and Claims Data.

Patients with diabetes face elevated hip fracture risk and post-fracture mortality compared to patients without diabetes. Early identification of high-risk individuals could enable targeted preventive interventions; however, existing models, such as the Fracture Risk Assessment Tool (FRAX), are known to underperform in patients with diabetes. For example, studies of FRAX have reported lower area under the curve (AUC) in patients with diabetes of 0.75-0.80, indicating moderate discrimination. While people have applied machine learning approaches to improve hip fracture prediction, validated models for patients with diabetes remain scarce. Most high-performing models rely on imaging data (e.g., DXA, radiographs) that may not be available.
We developed machine learning models using administrative data to predict hip fracture risk in patients with diabetes. Using the MarketScan commercial claims database (2003–2020), we selected patients aged ≥20 years with continuous enrollment for ≥1 year and type 1 or type 2 diabetes identified via ICD-9/ICD-10 codes (n=168,236). Diagnostic and procedural codes were mapped to phecodes. Phecodes aggregate related ICD-9 and ICD-10 codes into clinically meaningful phenotype categories, reducing redundancy and standardizing features across coding systems. Given the low event rate, we applied over-sampling to balance hip fracture cases and controls in the training set. We evaluated categorical boosting (CatBoost), extreme gradient-boosting decision trees (XGBoost), and convolutional and residual neural network architectures. We performed external validation on a de-identified cohort with diabetes from the University of Chicago (n=13,551). This cohort demonstrated a higher hip fracture prevalence (4.21% vs. 3.61%) and a greater comorbidity burden than the MarketScan cohort.
Each machine learning architecture underwent hyperparameter-tuning with the best-performing models validated on the University of Chicago cohort. Models achieved a mean area under the curve (AUC) of 0.94 on internal validation (MarketScan) and 0.91 on external validation (University of Chicago). The most important variables in the models were age, screening for malignant neoplasms of the skin, osteoarthrosis, injury, anemia, etc. Models predicted hip fracture occurrence during the observation period; future work will evaluate fixed time horizons (e.g., 1-year, 5-year risk). These models show that administrative claims data can predict hip fracture risk in patients with diabetes without requiring imaging, suggesting potential utility for the identification of high-risk individuals during routine clinic visits.

*Unless otherwise noted, all abstracts presented at ENDO must not be released to the press or the public until the date and time of presentation. For oral presentations, the abstracts are embargoed until the session begins. The Endocrine Society reserves the right to lift the embargo on specific abstracts that are selected for promotion prior to or during ENDO.*