Using Machine Learning to Predict Hip Fracture Risk from Dual-Energy X-Ray Absorptiometry Images and Health Factors

Abstract

Osteoporosis is a highly prevalent skeletal disease that greatly increases the risk of fragility fractures and affects millions of older adults worldwide. Hip fractures are especially dangerous, often resulting in long-term disability, increased need for institutional care, and increased healthcare costs. Current methods for diagnosing osteoporosis and subsequent fracture risk, such as bone mineral density (BMD) T-score from dual-energy x-ray absorptiometry (DXA) imaging and the Fracture Risk Assessment Tool (FRAX) have limitations in their ability to accurately predict fracture risk. Developments in machine learning and image processing tools have shown promise for improved fracture risk prediction. This work builds on these advances by integrating DXA imaging data with longitudinal clinical information to develop and evaluate predictive models aimed at improving fracture risk prediction.

To evaluate the independent predictive ability of DXA images, a simple feed-forward neural network trained on DXA images reduced using principal component analysis (PCA) was developed. This model demonstrated that images alone contain predictive information for fracture risk; however, performance outcomes improved when clinical risk factors (CRFs) were incorporated in a multimodal feed-forward neural network approach. SHapley Additive exPlanations analysis revealed that imaging features contributed most strongly to the model’s predictions while age and body mass index (BMI) had comparatively smaller contributions. Both models achieved performance metrics within the range of values typically reported for FRAX-based assessments. Next, deep learning and transfer learning models were developed to assess how model performance was impacted by the addition of spatial pattern recognition within the images. Deep learning achieved moderate predictive ability, with performance similar to FRAX, while transfer learning models struggled to learn meaningful patterns from the available data.

The purpose of this work was to improve fracture risk prediction by integrating DXA imaging data with clinical health information. The results of this work can inform the development of clinically relevant fracture risk prediction tools that support early intervention for at-risk individuals.