Seoul National University Hospital Develops Deep Learning Model to Predict Hip Fracture Re-fracture Risk with High Accuracy, Outperforms Existing Tools

Seoul National University Hospital analyzed CT scans of 1,480 hip fracture patients and designed a deep learning model to predict short-term risk of re-fracture
Excellent predictive accuracy (AUROC) of 0.74…Outperforms existing fracture prediction tools

[이데일리 이순용 기자]A method to easily predict the risk of re-fracture in hip fracture patients is presented. A deep learning prediction model developed by a domestic research team predicted the short-term risk of hip joint re-fracture within 5 years with high accuracy. It is expected that this will help establish existing management and treatment strategies for hip fractures.

The research team, including Professor Young-gon Kim from the Department of Integrative Medicine at Seoul National University Hospital, Researcher Lee Isaac Kim from the Biomedical Research Institute, and the -Professor Seong-hye Kong from the Department of Endocrinology and Metabolism at Seoul Bundang National University. Hospital, developed a short-term re-fracture risk prediction model based on CT images of 1,480 hip fracture patients and confirmed the accuracy mentioned on the 17th.

It is known that patients with hip fracture are at high risk of re-fracture, and that re-fracture occurs on average 2 to 4.3 years after the first fracture. Therefore, it is important to predict the short-term risk of re-fracture and monitor high-risk groups, but existing fracture prediction tools (FRAX, etc.) have limitations in short-term forecasting, so a new forecasting method is needed.

The research team focused on ‘hip joint CT images’, which can determine muscle and bone composition, to develop a short-term re-fracture risk prediction model. From January 2004 to December 2020, the hip joint CT images of 1,012 patients who visited the hospital with a fracture were reconstructed to create frontal, lateral and cross-sectional images. After extracting the features of each image, we designed an ensemble deep learning model that combines them and expresses the degree of freedom of refracture risk (probability of no refracture) in the form of a survival curve .

In addition, all patients were analyzed and ‘reference values’ were set for each time point that passed after the CT scan. If the risk tolerance is lower than this reference value, the probability of re-break increases. Therefore, by comparing the reference value curve with the patient’s survival curve, the time when the survival curve becomes lower than the reference value curve can be predicted as the time in which the event re-break.

Structural diagram of the deep learning model developed by the research team. Frontal, lateral and cross-sectional radiographs of the hip joint are examined and the degree of freedom for the risk of recurrence is calculated by combining them and expressed in the form of a survival curve. The point at which the patient’s survival curve (blue line) is lower than the reference value curve (red line) can be predicted as the point where relapse occurs.

In addition, as a result of testing the performance of 468 hip fracture patients, the prediction accuracy (AUROC) of the ensemble deep learning model that predicted short-term recurrence was high at approximately 0.74 (prediction of re-fracture within 2, 3, and 5 years). Accuracy 0.74, 0.74, 0.73, respectively) The closer the AUROC is to 1, the better the prediction performance.

This was a better performance than the FRAX prediction tool which is based on clinical information and bone density. The accuracy of FRAX for predicting re-fracture within 2, 3, and 5 years was 0.58, 0.64, and 0.70, respectively. Based on these findings, the research team emphasized that the CT-based deep learning prediction model can accurately predict short-term recurrence risk for less than 5 years, and explained that this could be used to establish management and treatment strategies for experienced patients. did a hip fracture.

Professor Kim Young-gon (first author) said, “Using the deep learning model developed by the research team will actively help identify high-risk groups for re-fracture,” and added, “For these high-risk groups, we provide prescriptions for osteoporosis, ongoing monitoring, and early rehabilitation.

This study was published in January in Radiology, an international journal in the field of radiology.