OA22 Machine learning and computer vision of bone microarchitecture can improve the fracture risk prediction provided by DXA and clinical risk factors

Abstract

Background/Aims

High-resolution peripheral quantitative computed tomography (HR-pQCT) scanning provides such detailed, 3-dimensional reconstructions of the skeleton that the images have been called ‘a virtual bone biopsy’. Traditional analysis of the images results in a multitude of cortical and trabecular parameters which would be potentially cumbersome to interpret for clinicians compared to user-friendly tools such as FRAX®. A computer vision approach, where the entire scan is ‘read’ by a computer algorithm to ascertain fracture risk, would be far simpler. Thus, we investigated whether a computer vision and machine learning technique could improve the current methods of assessing fracture risk.

Methods

This study was nested in the Hertfordshire Cohort Study, a group of community-dwelling older adults. Participants attended research visits at which height and weight were measured; fracture history was determined via self-report and vertebral fracture assessment. Bone microarchitecture was assessed via HR-pQCT scans of the non-dominant distal tibia (Scanco XtremeCT) and bone mineral density measurement and lateral vertebral assessment were performed using dual X-ray absorptiometry (DXA) (Lunar Prodigy Advanced). Images were cropped and pre-processed and texture analysis was performed using a 3-dimensional local binary patterns method. These analyses, together with age, sex, height, weight, BMI, and dietary calcium, were used in the random-forest classification algorithm. Receiver operating characteristic (ROC) analysis was used to compare fracture risk identification methods.

Results

Overall, 247 males and 149 females were included in this study with a mean age of approximately 76 years. Using clinical risk factors alone gave an area under the curve (AUC) of 0.70 (95% CI: 0.56-0.84), which improved to 0.71 (0.57-0.85) with the addition of DXA-measured BMD. The addition of the machine learning classifier to clinical risk factors and DXA-measured BMD lead to an improved AUC of 0.90 (0.83-0.96) with a sensitivity of 0.83 and specificity of 0.74.

Conclusion

The results of this preliminary work demonstrate that using a 3-dimensional computer vision method to HR-pQCT scanning can enhance the identification of those at risk of fracture beyond that afforded by clinical risk factors and DXA-measured BMD. This approach has the potential to make the information offered by HR-pQCT more accessible and applicable to healthcare professionals in the clinic if the technology becomes more widely available. Whilst these findings require replication in other cohorts, they are an early indicator that the application of a machine learning technique to bone microarchitecture could improve fracture prediction and osteoporosis care.

Disclosure

N.R. Fuggle: None. S. Lu: None. M. Ó Breasail: None. L.D. Westbury: None. K.A. Ward: None. E. Dennison: None. S. Mahmoodi: None. M. Niranjan: None. C. Cooper: None.