A CT-based Deep Learning Model for Predicting Subsequent Fracture Risk in Patients with Hip Fracture
Abstract
Background
Patients have the highest risk of subsequent fractures in the first few years after an initial fracture, yet models to predict short-term subsequent risk have not been developed.
Purpose
To develop and validate a deep learning prediction model for subsequent fracture risk using digitally reconstructed radiographs from hip CT in patients with recent hip fractures.
Materials and Methods
This retrospective study included adult patients who underwent three-dimensional hip CT due to a fracture from January 2004 to December 2020. Two-dimensional frontal, lateral, and axial digitally reconstructed radiographs were generated and assembled to construct an ensemble model. DenseNet modules were used to calculate risk probability based on extracted image features and fracture-free probability plots were output. Model performance was assessed using the C index and area under the receiver operating characteristic curve (AUC) and compared with other models using the paired t test.
Results
The training and validation set included 1012 patients (mean age, 74.5 years ± 13.3 [SD]; 706 female, 113 subsequent fracture) and the test set included 468 patients (mean age, 75.9 years ± 14.0; 335 female, 22 subsequent fractures). In the test set, the ensemble model had a higher C index (0.73) for predicting subsequent fractures than that of other image-based models (C index range, 0.59–0.70 for five of six models; P value range, < .001 to < .05). The ensemble model achieved AUCs of 0.74, 0.74, and 0.73 at the 2-, 3-, and 5-year follow-ups, respectively; higher than that of most other image-based models at 2 years (AUC range, 0.57–0.71 for five of six models; P value range, < .001 to < .05) and 3 years (AUC range, 0.55–0.72 for four of six models; P value range, < .001 to < .05). Moreover, the AUCs achieved by the ensemble model were higher than that of a clinical model that included known risk factors (2-, 3-, and 5-year AUCs of 0.58, 0.64, and 0.70, respectively; P < .001 for all).
Conclusion
In patients with recent hip fractures, the ensemble deep learning model using digital reconstructed radiographs from hip CT showed good performance for predicting subsequent fractures in the short term.
© RSNA, 2024
Supplemental material is available for this article.
See also the editorial by Li and Jaremko in this issue.
Summary
A deep learning model using digitally reconstructed radiographs from three-dimensional hip CT showed good performance for predicting short-term subsequent fractures (<5 years) in patients with a recent hip fracture.
Key Results
■ This retrospective study used 1480 digitally reconstructed radiographs from hip CT in patients with a recent hip fracture to train a deep learning model to predict subsequent fracture risk.
■ The deep learning model showed better performance at the 2- and 3-year follow-ups (area under the receiver operating characteristic curve [AUC], 0.74 for both) than most other image-based models (2-year AUC range, 0.57–0.71; 3-year AUC range, 0.55–0.72; P value range, < .001 to < .05) and a clinical model that included known risk factors (2- and 3-year AUCs, 0.58 and 0.64; P < .001 for both).
Introduction
Osteoporotic fractures are associated with increased mortality and disability (1,2), which increases further as patients age (3). Patients who experience fracture are also more likely to have subsequent fractures, with the risk being highest in the first few years after the initial fracture (4,5). However, studies aiming to identify patients most likely to experience subsequent fractures among those with a first fracture are scarce. Tools have been developed to predict osteoporotic fractures in the long term, 5 or 10 years, such as the Fracture Risk Assessment Tool (FRAX) (6) and Garvan Fracture Risk Calculator (7). These fracture risk prediction tools incorporate previous fracture history (6,7) but cannot predict subsequent fractures in patients who have recently experienced a first fracture.
Developing a prediction model for short-term subsequent fracture risk is important because it would identify patients at the highest risk and aid in determining appropriate treatment strategies. For instance, patients identified as being at a higher risk of subsequent fractures could benefit from comprehensive monitoring for initiating osteoporosis medication and ensuring compliance, along with frequent follow-up within the first few years (8). Additionally, early rehabilitation could be strongly recommended as a preventive measure against falls, which are known to increase the risk of subsequent fractures (9). Nevertheless, developing a prediction model for subsequent fractures has been challenging as the relationship between bone mineral density (BMD) and the risk of subsequent fractures may suggest that BMD is not the sole factor explaining risk (10,11).
Recently, techniques for medical image analysis have undergone rapid advancement, starting with the rise of the convolutional neural network (CNN) (12). CNNs enable highly representative and layered hierarchical image features, which can successfully classify medical images according to diagnosis and prognosis. Previous studies have attempted to use CNNs and various medical images to identify patients with high fracture risk (13–16). However, previous studies focused on initial fracture events rather than subsequent osteoporotic fractures (13–16). As muscle mass and quality could be critical contributing factors to subsequent osteoporotic fractures in those with recent fractures (17,18), CT could be a valuable method to assess muscle and bone. Nevertheless, a prediction model for subsequent fractures using CT images has not been developed. Therefore, the aim of this study was to develop and validate a deep learning prediction model for subsequent fracture risk using digitally reconstructed radiographs from hip CT in patients with recent hip fractures.
Materials and Methods
Study Sample
This retrospective study included longitudinal data from adult patients who underwent three-dimensional (3D) hip CT due to hip fracture at Seoul National University Bundang Hospital from January 2004 to December 2020. To ensure data accuracy, patients were excluded if they did not experience a fracture within 3 months of the 3D hip CT scan. Patients with a follow-up duration less than 2 years, poor image quality, and those with subtrochanteric fractures were also excluded (Fig 1). The local institutional review board approved the study protocol (B2104677402), and the requirement for written informed consent was waived due to the retrospective nature of the study. The study was conducted following the World Medical Association Declaration of Helsinki and its ethical principles for medical research.
Figure 1: Flowchart shows study inclusion and exclusion. 3D = three-dimensional.
Clinical Variables
The clinical variables included in this study were the same as the variables included in the FRAX model, which are femoral neck BMD T score, age, sex, body mass index (calculated as weight in kilograms divided by height in meters squared), fracture history, parental hip fracture history, secondary osteoporosis, current smoking and alcohol consumption status, corticosteroid use, and the presence of rheumatoid arthritis. Measurement details of clinical parameters are included in Appendix S1.
Imaging Protocols
Lumbar spine, femoral neck, and total hip BMDs (in grams per square centimeter) were measured using dual x-ray absorptiometry (Discovery W; Hologic) and analyzed according to manufacturer instructions. Images were obtained using CT scanners with 64 detector rows (Brilliance; Philips). All patients were positioned supine and scanned with a length of approximately 400 mm, extending from 10 mm above the iliac crest to the middle femoral shaft. Measurement details are included in Appendix S1.
Image Processing and the Deep Learning Model
A deep learning–based convolutional neural network model was developed to predict subsequent fractures using digitally reconstructed radiographs generated from CT. Although each digitally reconstructed radiograph is two-dimensional (2D), the prediction score was calculated using the frontal, lateral, and axial images. Because the model uses multiple 2D images from various directions and, thus, could be in the middle of 2D and 3D, it was called a 2.5-dimensional approach. Figure 2 provides an example of the digitally reconstructed radiographs.
Figure 2: (A) Diagram shows the image processing procedure. Three-dimensional volume is displayed by the average of all intensities from the x-ray source to the projection plane forming the digitally reconstructed radiograph (DRR) and cropped to a square to fit the body area based on the frontal view. (B) Diagram shows image processing examples for frontal (top), lateral (middle), and axial (bottom) digitally reconstructed radiographs.
A DenseNet-based model (19), pretrained on the ImageNet data set (20), was trained to extract imaging features for each digitally reconstructed radiographic view. Deep logistic hazard loss function (21) was used to estimate survival function. The model architecture is depicted in Figure 3 and additional details are included in Appendix S1. The codes are publicly available at https://github.com/Medical-Vision-Lab/SubsequentFracturePrediction.
Figure 3: Diagram shows the architecture of the 2.5-dimensional ensemble model. DenseNet-based modules calculate the risk probability of fracture based on the frontal, lateral, and axial digitally reconstructed radiographs and assembles this into a decision aggregator. The final output is a subsequent fracture–free probability plot in the form of survival function.
Models for Clinical Data
The clinical models were trained using the XGBoost algorithm (21), and the loss functions for learning accelerated failure time was implemented to increase the support for survival modeling (22). Model A was trained with age and sex; model B, additionally, with body mass index, smoking and alcohol consumption status, steroid use, presence of rheumatoid arthritis, and secondary osteoporosis; and model C, additionally, with hip fracture type, femoral neck BMD, and use of antiosteoporosis medication. Predictors were prioritized based on a combination of domain knowledge and literature review, especially based on factors included in the well-established FRAX model.
Definition of Subsequent Fracture
The primary outcome was subsequent osteoporotic fracture, defined as fracture due to low-level trauma such as a fall from standing height or less, including vertebral, hip, humerus, and wrist fractures, occurring greater than or equal to 3 months after an initial osteoporotic fracture. Appendix S1 provides additional details.
Performance Evaluation and Statistical Analysis
Concerning baseline characteristics, continuous parameters were presented as means with SDs and categorical data as proportions. Comparisons between groups were analyzed using the Student t test for continuous variables and the χ2 test for categorical variables. The area under the receiver operating characteristic curve (AUC), area under the precision-recall curve, and concordance index (C index) were calculated to compare models. All models were internally evaluated using 10-fold cross-validation (training and validation set) and tested with a temporally separated data set (test set). An optimized threshold using the Youden J statistic was determined from the validation data set and applied to the test data set. The paired t test was used to compare the proposed 2.5-dimensional ensemble model with other models. P < .05 was considered indicative of a statistically significant difference. Calibration analysis is detailed in Appendix S1. PyTorch, Scikit-learn, pycox (23), and XGBSE (21) libraries from Python (version 3.6.9, Python Software Foundation) were used for the analyses performed by two authors (Y.K. and B.W.K).
Results
Baseline Patient Characteristics
A total of 10 381 patients were initially considered for inclusion. Patients were excluded if they did not experience a fracture within 3 months of the initial 3D hip CT scan (n = 7041), had a follow-up duration of less than 2 years (n = 1430), had scans with poor image quality (n = 381), or had subtrochanteric fractures (n = 49) (Fig 1). Thus, CT images from 1480 individuals (mean age, 75.5 years ± 13.3 [SD]; 1041 female [70.3%], 439 male [29.7%]) were included in the final analysis.
During the mean follow-up duration of 3.4 years ± 1.1, 135 of 1480 patients (9.1%) experienced subsequent osteoporotic fractures (hip, 86 [63.7%]; vertebral, 45 [33.3%]; other nonvertebral fractures, four [3.0%]). Patients who experienced subsequent fractures were older (mean age, 77.9 years ± 8.4 vs 74.6 years ± 13.9; P < .001), more likely to be female (107 of 135 [79.3%] vs 934 of 1345 [69.4%], P = .02), more likely to use steroids (14 of 135 [10.4%] vs 60 of 1345 [4.5%], P = .01), and more likely to have secondary osteoporosis (24 of 135 [17.8%] vs 140 of 1345 [10.4%], P = .01) than those who did not (Table 1). No evidence of a difference in antiosteoporosis medication use was observed between patients who did and did not experience subsequent fractures (P = .082). Among the 907 of 1480 patients (61.3%) who used antiosteoporosis medication, 714 of 1480 (48.2%) used bisphosphonates, 119 of 1480 (8.0%) used teriparatide, 45 of 1480 (3.0%) used denosumab, and 29 of 1480 (2.0%) used selective estrogen receptor modulators.
 |
The risk of major osteoporotic and hip fracture calculated using the FRAX for major osteoporotic fracture were higher in patients who experienced subsequent fractures than in those who did not (mean FRAX score for major osteoporotic fracture, 8.2 ± 4.1 vs 7.0 ± 4.0; P < .01). No evidence of a difference between fracture type, body mass index, or BMD of the lumbar spine, femur neck, and total hip were observed between groups (P value range, .12–.47).
Images obtained in 1012 patients before 2014 were included in the development set for model training and validation, and images obtained in 468 patients after 2015 were used for the test set. When comparing the data sets, patients in the test set had a lower number of subsequent fractures (22 of 468 [4.7%] vs 113 of 1012 [11.2%], P < .01) and were less likely to have secondary osteoporosis (36 of 468 [7.7%] vs 128 of 1012 [12.6%], P = .01) than those in the training and validation set. However, no evidence of a difference between other baseline characteristics, including age, sex, body mass index, smoking or alcohol consumption status, and BMD, were observed between groups (P value range, .06–.60). No patients in the training and validation data set used denosumab, while 45 patients in the test set did (Table S1).
Comparing Performance of the 2.5-dimensional Ensemble Model to Other Image-based Models
As shown in Table 2, in the training and validation data set, the 2.5-dimensional ensemble model had a higher mean C index (0.73 [95% CI: 0.67, 0.79]) for predicting subsequent fractures than models using hip anteroposterior radiographs (0.66 [95% CI: 0.61, 0.72]); scout film from CT (0.67 [95% CI: 0.62, 0.72]); frontal (0.70 [95% CI: 0.64, 0.75]), axial (0.70 [95% CI: 0.64, 0.75]), and lateral (0.66 [95% CI: 0.60, 0.72]) digitally reconstructed radiographs; and 3D volume (0.60 [95% CI: 0.54, 0.67]). Also, for 2-, 3-, and 5-year fracture prediction, the 2.5-dimensional ensemble model showed mean AUC values of 0.78 (95% CI: 0.71, 0.85), 0.76 (95% CI: 0.67, 0.85), and 0.77 (95% CI: 0.70, 0.84), which were higher than all other models.
 |
In the test set, the 2.5-dimensional ensemble model yielded a C index of 0.73, which was higher than that of all other models except the lateral digitally reconstructed radiographic model. The 2.5-dimensional ensemble model achieved an AUC of 0.74 for predicting subsequent fractures at the 2-year follow-up, which was higher than that of all models except scout film from CT. At the 3-year follow up, the AUC of the 2.5-dimensional model was also higher (0.74), except compared with that of the lateral digitally reconstructed radiographic model. At the 5-year follow-up, the AUC achieved by the 2.5-dimensional model was higher than that of the hip radiography model (0.73 vs 0.57; P < .001).
Examples of model predictions are presented in Figure 4. The model was built as a web-based interface (Fig S1), and the heat map images in Figures 4, S2, and S3 were developed using gradient-weighted class activation mapping (22). Receiver operating characteristic curves are shown in Figures 5 and S4, and areas under the precision-recall curve are presented in Table S2.
Figure 4: An example case from the test set. (A) Left to right: Frontal, lateral, and axial digitally reconstructed radiographs in a 76-year-old male patient show initial intertrochanteric fracture (input images for the ensemble model); no subsequent fracture was seen over 1835 days of follow-up. (B) Gradient-weighted class activation maps from the deep learning model show model prediction is influenced by the femur neck area (red) in each corresponding radiographic view. (C) Subsequent fracture–free probability plot in the form of survival function as predicted by the 2.5-dimensional ensemble model shows the patient is likely to experience subsequent fracture at points on the graph where the case line (blue) is lower than the reference line (red). The case line is generated directly from outputs of the developed model, and the reference line is the optimized threshold for each time point. Predicted risk of 2-, 3-, and 5-year fracture were 10.8%, 20.7%, and 19.4%, respectively; that is, the patient is unlikely to experience subsequent fracture.
Figure 5: Receiver operating characteristic curves in the test set show (A) 2-, (B) 3-, and (C) 5-year prediction of subsequent fracture risk, whereby the 2.5-dimensional (2.5D) ensemble model had a higher area under the receiver operating characteristic curve than the models using two-dimensional (2D) frontal digitally reconstructed radiographs (DRR) and three-dimensional (3D) volume.
As the rate of denosumab use differed between the training and validation data set and test data set, the false-positive rate between patients who did or did not have denosumab was evaluated. False-positive rates were higher in patients who used denosumab than in those who did not (0.44 vs 0.26 at 2 years, 0.67 vs 0.47 at 3 years).
Performance of the CT-based 2.5-dimensional Ensemble Model Versus Clinical Models
In the training and validation set, the 2.5-dimensional ensemble model achieved a C index of 0.73 (95% CI: 0.67, 0.79) for subsequent fracture prediction and AUCs of 0.78, 0.76, and 0.77 for the 2-, 3-, and 5-year follow-ups, respectively (Table 3). All these performance metrics were higher than those achieved by clinical model C, which included known risk factors, (C index, 0.60 [95% CI: 0.55, 0.65]; AUC at 2 years, 0.66 [95% CI: 0.57, 0.75]; AUC at 3 years, 0.59 [95% CI: 0.52, 0.67]; AUC at 5 years, 0.65 [95% CI: 0.58, 0.72]). In the test set, the 2.5-dimensional ensemble model also showed a higher C index and AUCs at 2, 3, and 5 years than clinical model C (C index, 0.73 [95% CI: 0.62, 0.90], 0.62 [95% CI: 0.43, 0.78]; AUC at 2 years, 0.74 [95% CI: 0.60, 0.88], 0.58 [95% CI: 0.43, 0.73]; AUC at 3 years, 0.74 [95% CI: 0.61, 0.87], 0.64 [95% CI: 0.54, 0.75]; AUC at 5 years, 0.73 [95% CI: 0.61, 0.86], 0.70 [95% CI: 0.58, 0.81], respectively).
 |
FRAX achieved a C index of 0.56 and 0.60 with and without BMD, respectively, for predicting major osteoporotic fracture in the test set. For predicting hip fracture in the test set, FRAX achieved C indexes of 0.57 and 0.58 with and without BMD, respectively. FRAX could not be directly compared with the ensemble model due to the difference in follow-up duration. Receiver operating characteristic curves are shown in Figure 6 and areas under the precision-recall curve are presented in Table S3. Also, receiver operating characteristic curves for all image models in the test set are presented in Figure S5.
Figure 6: Receiver operating characteristic curves in the test set show (A) 2-, (B) 3-, and (C) 5-year prediction of subsequent fracture risk, whereby the 2.5-dimensional ensemble model had a higher area under the receiver operating characteristic curve than the Fracture Risk Assessment Tool (FRAX) with bone mineral density for predicting major osteoporotic fracture and clinical model C.
An optimal threshold using the Youden J statistic was determined from the validation set. When applying the optimal thresholds of 0.04, 0.07, and 0.13, the sensitivities were 0.84, 0.91, and 0.84 and specificities were 0.84, 0.70, and 0.74 at the 2-, 3-, and 5-year follow-ups, respectively, for the 2.5-dimensional ensemble model. When applying the optimal thresholds in the test set, the sensitivities were 0.62, 0.87, and 0.71 and specificities were 0.72, 0.51, and 0.60 at the 2-, 3-, and 5-year follow-ups, which showed that the model consistently had high sensitivity in the test set, especially for prediction at 3 and 5 years.
Discussion
Patients have the highest risk of subsequent fractures in the first few years after initial fractures, yet models to predict short-term (<5 years) subsequent fracture risk have not been developed. In this study, we used digitally reconstructed radiographs obtained from three-dimensional hip CT in 1480 patients with recent hip fracture to develop and validate a deep learning prediction model for short-term subsequent fracture risk. In the test set (n = 468), the deep learning model showed better performance at the 2- and 3-year follow-ups (area under the receiver operating characteristic curve [AUC], 0.74 for both) than most other image-based models (2-year AUC range, 0.57–0.71; 3-year AUC range, 0.55–0.72; P value range, < .001 to < .05) and a clinical model that included known risk factors (2-year AUC, 0.58; 3-year AUC, 0.64; P < .001 for both).
Recently, a model predicting subsequent vertebral fractures using CT images in patients with a vertebral fracture has been reported (24). Using CT images from 103 patients with vertebral fractures and 70 age-matched patients with osteoporosis, the model had an accuracy of 0.84 and AUC of 0.88. Although it was not externally validated, it was encouraging due to its high performance, implying that images of the bone and adjacent muscle could help identify patients at high risk for subsequent fractures (24). Information from muscle images, such as area and degree of fat infiltration in the muscle, is correlated with skeletal stability and risk of fractures (25–27). This implies that information from muscles supporting and communicating with the bone could improve the model, in addition to bone density and quality (18), which can be readily obtained from the images. Although heat maps generated by our model appear to focus on areas that include both bone and muscle, in actuality these may be visualizations of the regions the model considered important to predict subsequent fractures rather than an accurate depiction of attention localization.
In the current study, FRAX showed relatively low performance for predicting fracture compared with previously published studies. This may be due to differences in study samples and design. Previous studies reported the performance of the FRAX tool to predict fracture for both first and subsequent fractures, while we specifically focused on predicting subsequent fractures in patients with very high fracture risk. Furthermore, due to the discrepancy in follow-up duration, our study aimed to investigate events that happened within 5 years, while FRAX presents 10-year fracture probabilities.
While Hu et al (24) have reported a prediction model of osteoporotic vertebral fractures with good performance, our study is the first to develop a subsequent fracture prediction model in patients with recent hip fracture; the AUCs we obtained were higher than that of clinical models. A strength of our study is that our model was tested on a distinct and independent temporal data set. It also showed that a model using images of various hip CT angles performed better compared with models with frontal cross-sectional images, such as anteroposterior radiographs, suggesting the potential use of the 2.5-dimensional model with other CT images.
Our study had several limitations. First, it used a specific CT protocol (3D hip CT), which is the main protocol used in our institution, possibly reducing the generalizability of the model. Second, it was difficult to explain what factors influenced the prediction of subsequent fractures. Although we checked the heat map images and found that the model mainly focused on the femur neck and shaft areas, the focus range varied widely among patients. Third, although we presented examples of probabilities of subsequent fractures in individual cases, the model still had uncertainties in reliability, which needs further validation studies. Finally, the proportion of patients who used denosumab was different between the training and validation set and the test data set, such that patients who used denosumab were only included in the test set. Because the false-positive rate of the 2.5-dimensional digitally reconstructed radiographic model was higher in patients who used denosumab than in those who did not, the presence of patients using denosumab only in the test set may have attenuated the performance.
In conclusion, a deep learning model using digitally reconstructed radiographs from three-dimensional hip CT showed good performance for predicting subsequent fractures in the short term (<5 years) in patients with recent hip fracture. Using a deep learning model to predict subsequent fracture risk from initial hip CT imaging may help identify patients at high risk who would benefit from proactive measures, such as thorough monitoring of the initiation and compliance of osteoporosis medication for the first few years and early rehabilitation to mitigate falls.
Author Contributions
Author contributions: Guarantors of integrity of entire study, S.H.K., Y.K.L., S.W.K., C.S.S.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, Y.K., Y.G.K., J.W.P., S.H.K., Y.K.L., S.W.K.; clinical studies, S.H.K., J.H.K., Y.K.L., S.W.K.; experimental studies, Y.K., Y.G.K., B.W.K., Y.S., S.H.K.; statistical analysis, Y.K., Y.G.K., B.W.K., Y.S., S.H.K.; and manuscript editing, Y.K., Y.G.K., J.W.P., B.W.K., S.H.K., J.H.K., Y.K.L., S.W.K., C.S.S.
* Y.K. and Y.G.K. contributed equally to this work.
Supported by the National Research Foundation of Korea (2021R1A2C200341012) and Seoul National University Bundang Hospital (08-2023-0085, 02-2021-0019).
References
- 1. . Health service use pathways associated with recovery of quality of life at 12-months for individual fracture sites: Analyses of the International Costs and Utilities Related to Osteoporotic fractures Study (ICUROS). Bone 2021;144:115805.
- 2. . Osteoporosis and Osteoporotic Fracture Fact Sheet in Korea. J Bone Metab 2020;27(4):281–290.
- 3. . An estimate of the worldwide prevalence and disability associated with osteoporotic fractures. Osteoporos Int 2006;17(12):1726–1733.
- 4. . Characteristics of recurrent fractures. Osteoporos Int 2018;29(8):1747–1757.
- 5. . The effect on subsequent fracture risk of age, sex, and prior fracture site by recency of prior fracture. Osteoporos Int 2021;32(8):1547–1555.
- 6. . Systematic review of major osteoporotic fracture to hip fracture incidence rate ratios worldwide: implications for Fracture Risk Assessment Tool (FRAX)-derived estimates. J Bone Miner Res 2021;36(10):1942–1956.
- 7. . Predictive performance of the Garvan Fracture Risk Calculator: a registry-based cohort study. Osteoporos Int 2022;33(3):541–548.
- 8. . Recommendations on the post-acute management of the osteoporotic fracture: patients with “very-high” re-fracture risk. J Orthop Translat 2022;37:94–99.
- 9. . A multidisciplinary intervention to prevent subsequent falls and health service use following fall-related paramedic care: a randomised controlled trial. Age Ageing 2017;46(2):200–207.
- 10. . A meta-analysis of previous fracture and subsequent fracture risk. Bone 2004;35(2):375–382.
- 11. . Risk of subsequent fractures and mortality in elderly women and men with fragility fractures with and without osteoporotic bone density: the Dubbo Osteoporosis Epidemiology Study. J Bone Miner Res 2015;30(4):637–646.
- 12. . Medical Image Analysis using Deep Convolutional Neural Networks: CNN Architectures and Transfer Learning. In:
2020 International Conference on Inventive Computation Technologies (ICICT) . IEEE, 2020; 175–180. - 13. . Development of a Spine X-Ray-Based Fracture Prediction Model Using a Deep Learning Algorithm. Endocrinol Metab (Seoul) 2022;37(4):674–683.
- 14. . Identification of Vertebral Fractures by Convolutional Neural Networks to Predict Nonvertebral and Hip Fractures: A Registry-based Cohort Study of Dual X-ray Absorptiometry. Radiology 2019;293(2):405–411.
- 15. . Application of the Trabecular Bone Score in Clinical Practice. J Bone Metab 2021;28(2):101–113.
- 16. . Applications of Machine Learning in Bone and Mineral Research. Endocrinol Metab (Seoul) 2021;36(5):928–937.
- 17. . Computed tomography-based paravertebral muscle density predicts subsequent vertebral fracture risks independently of bone mineral density in postmenopausal women following percutaneous vertebral augmentation. Aging Clin Exp Res 2022;34(11):2797–2805.
- 18. . Quality Matters as Much as Quantity of Skeletal Muscle: Clinical Implications of Myosteatosis in Cardiometabolic Health. Endocrinol Metab (Seoul) 2021;36(6):1161–1174.
- 19. . Densely Connected Convolutional Networks. In:
2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE, 2017; 2261–2269. - 20. . ImageNet: A large-scale hierarchical image database. In:
2009 IEEE Conference on Computer Vision and Pattern Recognition . IEEE, 2009; 248–255. - 21. . XGBoost survival Embeddings: improving statistical properties of XGBoost survival analysis implementation. https://github.com/loft-br/xgboost-survival-embeddings. Published 2021. Accessed November 8, 2023.
- 22. . Survival regression with accelerated failure time model in XGBoost. J Comput Graph Stat 2022;31(4):1292–1302.
- 23. . Time-to-Event Prediction with Neural Networks and Cox Regression. arXiv 1907.00825 [preprint] https://arxiv.org/abs/1907.00825. Published July 1, 2019. Accessed November 8, 2023.
- 24. . Prediction of subsequent osteoporotic vertebral compression fracture on CT radiography via deep learning. VIEW 2022;3(6):20220012.
- 25. . Cross-Sectional Area of the Lumbar Spine Trunk Muscle and Posterior Lumbar Interbody Fusion Rate: A Retrospective Study. Clin Spine Surg 2017;30(6):E798–E803.
- 26. . Impact of paravertebral muscle in thoracolumbar and lower lumbar regions on outcomes following osteoporotic vertebral fracture: a multicenter cohort study. Arch Osteoporos 2021;16(1):2.
- 27. . Changes of paraspinal muscles in postmenopausal osteoporotic spinal compression fractures: magnetic resonance imaging study. J Bone Metab 2013;20(2):75–81.
Article History
Received: Mar 12 2023Revision requested: Apr 17 2023
Revision received: Dec 4 2023
Accepted: Dec 20 2023
Published online: Jan 30 2024
