Bone and Mineral Metabolism

SUN-809 - Fracture Risk Prediction in Adults over 50 Year Old with Traditional and Machine Learning Models in Two Large Cohorts

Sunday, June 14, 2026

9:00 AM - 4:00 PM CT

Location: ENDOExpo; Poster Floor

Abstract Poster Presenter(s)

YL

Yi Liu, MD

Weill Cornell Medical College
New York, New York, United States

Senior Author/Primary Investigator(s)

YL

Yi Liu, MD

Weill Cornell Medical College
New York, New York, United States

Co-Author(s)

JQ

Jiahe Qian, MS

Weill Cornell Medical College
New York, New York, United States
HD

Hao Dai, PhD

Indiana University
Indianapolis, United States
KY

Kunyu Yu, BS

Weill Cornell Medical College
New York, United States
HD

Hexin Dong, PhD

Weill Cornell Medical College
New York, United States
JZ

Jiali Zhang, MS

Indiana University
Indianapolis, United States
JB

Jiang Bian, PhD

Indiana University
Indianapolis, United States

Senior Author/Primary Investigator(s)

YP

Yifan Peng, PhD

Weill Cornell Medical College
New York, United States

Co-Author(s)

YP

Yifan Peng, PhD

Weill Cornell Medical College
New York, United States

Introduction: Fracture prediction is central to osteoporosis management, yet the most commonly used tool, the Fracture Risk Assessment Tool (FRAX), has notable limitations. Machine learning (ML) has emerged as a promising technique for risk prediction, though its performance compared to traditional methods remains controversial.

Methods: We developed both traditional Cox models incorporating clinical risk factors, T-scores, and treatment history, as well as ML models to identify patients at risk for fragility fracture. A total of 11,522 patients aged over 50 years with at least one DXA scan were included in the Weill Cornell Medical College (WCMC) cohort between January 2012 and March 2023, with 859 patients sustaining fragility fractures. External validation was performed in a cohort from Indiana University (IU; n=9,180 with 852 fracture cases). Uno's C-index and time-dependent area under the receiver operating characteristic curve (AUC) were calculated to evaluate the performance of different ML models (random survival forest [RSF], eXtreme Gradient Boosting [XGBoost], and gradient boosting survival) compared to Cox regression and FRAX.

Results: Cox Model A, which incorporates comprehensive clinical risk factors, the lowest T-score among all sites, and treatment information, demonstrated the highest C-index for 2-year fragility fracture prediction (C-index 0.813, 95% confidence interval [CI] 0.806–0.820). ML models showed lower performance than Cox regression, with RSF demonstrating the best performance among the ML approaches (C-index 0.804, 95% CI 0.796–0.811). The 10-year fracture probability from FRAX showed the poorest discrimination among all three approaches (C-index 0.663, 95% CI 0.654-0.672). All pairwise comparisons between these methods were statistically significant (all p< 0.01). The most important variables for RSF, identified using Shapley additive explanation (SHAP) values, were prior fractures, age, lowest T-scores, comorbidities, and body mass index. External validation in the IU cohort generated similar results.

Discussion: Fracture prediction performance was significantly better from Cox regression model than ML models, with both substantially outperforming FRAX when incorporating comprehensive clinical variables, bone mineral density T-scores and treatment history.

*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.*