Deep Learning Model for Predicting Fracture Risk

Patients are at higher risk of subsequent fractures in the first few years following an initial fracture. Models to predict short-term subsequent risk have not been developed.

A deep-learning model using digital x-rays reconstructed from 3D hip CT images for predicting short-term subsequent fractures (<5 years) in patients with a recent hip fracture is promising, according to a study by a team led by Yisak Kim of Seoul National University Graduate School. This study could improve clinician care for their patients, and the findings of this Original Research on Musculoskeletal Imaging are published in Radiology.

Study Purpose

The primary purpose of this study was to develop and validate a deep-learning prediction model using digitally reconstructed radiographs from hip CT in patients (recent hip fractures) to predict subsequent fracture risk. The study included patients who underwent three-dimensional hip CT from January 2004 to December 2020 and generated two-dimensional frontal, lateral, and axial radiographs. These were assembled to construct an ensemble model. DenseNet modules calculated risk probability based on extracted image features and output were fracture-free probability plots. C index and AUC assessed model performance and compared with other models using the paired t-test.

Key Findings

• The training and validation set had 1012 patients with a mean age of 74.5 years, including 706 females and 113 subsequent fractures.
• The test set had 468 patients of mean age, 75 years, with 335 females and 22 subsequent fractures.
• In the test set, the ensemble model had a higher C index (0.73) than other image-based models for predicting subsequent fractures (C index range, 0.59–0.70 for five of six models).
• The ensemble model achieved AUCs of 0.74, 0.74, and 0.73 at the follow-ups (2-, 3-, and 5 years), respectively. This is higher than most other image-based models at two years (AUC range, 0.57–0.71 for five of six models) and three years (AUC range, 0.55–0.72 for four of six models).
• Moreover, the ensemble model achieved higher AUCs than the clinical model [included known risk factors (AUCs of 0.58, 0.64, and 0.70, respectively in follow-up)].

Concluding further, among those with recent hip fractures, the ensemble deep learning model using digitally reconstructed radiographs from hip CT demonstrated good performance for predicting subsequent fractures in the short term.

Reference

Kim, Y et al. A CT-based Deep Learning Model for Predicting Subsequent Fracture Risk in Patients with Hip Fracture. Radiology, 310(1).