Deep Learning Using High-Resolution Images of Forearm Predicts Fracture

Importance

Fragility fractures are a public health problem. Over 70% of women having fractures have osteopenia or normal BMD, but they remain unidentified and untreated because the definition of ‘osteoporosis’, a bone mineral density (BMD) T-Score ≤ -2.5SD, is often used to signal bone fragility.

Objective

As deep learning facilitates investigation of bone’s multi-level hierarchical structure and soft tissue, we tested whether this approach might better identify women at risk of fracture before fracture.

Design

We pooled data from three French and Swiss prospective population-based cohorts (OFELY, QUALYOR, GERICO) that collected clinical risk factors for fracture, areal BMD and distal radius measurements with high resolution peripheral quantitative tomography (HRpQCT). Using only three-dimensional images of the distal radius, ulna and soft tissue acquired by HRpQCT, an algorithm, a Structural Fragility Score-Artificial Intelligence (SFS-AI), was trained to distinguish 277 women having fractures from 1401 remaining fracture-free during 5 years and then was tested in a validation cohort of 422 women.

Setting

European postmenopausal women

Participants

We have studied postmenopausal women considered as representative of the general population, who were followed for a median 9.4 years in OFELY, 5.4 years in QUALYOR and 5.7 years in GERICO.

Main outcome and measure

All types of incident fragility fractures

Results

We used data from 2666 postmenopausal women, with age range of 42-94. In women ≥ 65 years having ‘All Fragility Fractures’ or ‘Major Fragility Fractures’, SFS-AI generated an AUC of 66-70%, sensitivities of 60-68% and specificity of 71%. Sensitivities were greater than achieved by the fracture risk assessment (FRAX) with BMD or BMD (6.7-26.7%) with lower specificities than these diagnostics (∼95%).

Conclusion and relevance

The SFS-AI is a holistic surrogate of fracture risk that pre-emptively identifies most women needing prompt treatment to avert a first fracture.

Key Points