Moderate evidence
More
Spine surgery study finds AI can flag cement leakage risk before vertebroplasty
Study title: Development of a Multimodal Model Combining Imaging and Clinical Data to Predict Bone Cement Leakage During Vertebroplasty for Fresh Thoracolumbar Compression Fractures
- Evidence type
- Controlled or validated comparison
- Sample
- n = 486
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2954cite or check it
Study title: Development of a Multimodal Model Combining Imaging and Clinical Data to Predict Bone Cement Leakage During Vertebroplasty for Fresh Thoracolumbar Compression Fractures
A hospital team in China built machine-learning models to predict whether bone cement would leak outside the vertebra during percutaneous vertebroplasty, a common spine fracture treatment. Using 486 treated vertebrae from 378 patients, they combined clinical data, CT imaging findings and radiomics features. Multimodal models reached AUCs of 0.902-0.939 on the internal test set.
Read the study abstract
Background: Bone cement leakage is a common complication of percutaneous vertebroplasty (PVP), but perioperative risk stratification remains challenging. Existing models based on single-source clinical or imaging features have limited performance. This study aimed to develop and validate multimodal artificial intelligence (AI) models integrating clinical, imaging, and radiomics features to predict bone cement leakage during PVP. Methods: This multicenter study included 486 treated vertebrae from 378 patients at…

The gain
In a study of 486 treated vertebrae from 378 patients, multimodal AI models predicted cement leakage during vertebroplasty with AUCs of 0.902-0.939 on internal testing and 0.834-0.925 in a prospective cohort, above single-source models.
In brief
Cement leakage occurred in 215 vertebrae (44.2%) and can cause serious problems such as neurological injury and pulmonary embolism. The models were also tested on 50 vertebrae from another hospital and 60 vertebrae from 50 prospectively enrolled patients, where AUCs ranged from 0.722-0.929 and 0.834-0.925.
The models predict leakage seen on postoperative imaging, not symptoms or complications. The authors say larger and more diverse populations are needed before wider clinical use, and that any injection-technique changes based on predicted risk need prospective validation.
Main points
- Leakage appeared in 215 of 486 treated vertebrae (44.2%) from 378 patients treated between January 2018 and June 2025 at three campuses of one hospital.
- Higher BMI (OR 1.11), larger cement volume (OR 6.18), cortical destruction (OR 6.39) and posterior wall involvement (OR 8.05) were independent risk factors.
- In the internal test set, the best single-modality models reached AUCs of 0.871 (clinical), 0.844 (imaging) and 0.806 (radiomics), versus 0.902-0.939 for multimodal models.
- XGBoost scored highest in internal (0.939) and external (0.929) validation; random forest scored highest in the prospective cohort (0.925).
- After Bonferroni correction, multimodal models beat their single-modality counterparts significantly in seven comparisons.
- SHAP analysis ranked cortical destruction, cement injection volume and posterior wall involvement as the most influential predictors.
- Removing three scanner-sensitive radiomics features lowered the SVM AUC by 0.022 to 0.036 across cohorts.
The full story
The retrospective development cohort split into 385 training vertebrae from 301 patients and 101 test vertebrae from 77 patients, with all vertebrae from one patient kept in the same partition. Leakage was judged on postoperative imaging by three spine surgeons and two musculoskeletal radiologists. Six algorithms were used to build clinical, imaging, radiomics and multimodal models.
In the internal test set, multimodal AUCs ranged from 0.902 to 0.939. In the external cohort of 50 vertebrae from 45 patients they ranged from 0.722 to 0.929, and in the prospective cohort of 60 vertebrae from 50 patients from 0.834 to 0.925. Patient-level aggregation gave AUCs of 0.865-0.912, and a test retaining one vertebra per patient gave mean AUCs of 0.889-0.918.
A separate analysis of 108 locked model configurations gave median AUCs of 0.899 internally, 0.919 externally and 0.860 prospectively. Decision curve analysis showed net benefit at threshold probabilities of roughly 0.10-0.60. The authors suggest high predicted risk could prompt slower injection or tighter cement volume control, but say this needs prospective validation.
Sources
1 on file- JOR SPINEdoi:10.1002/jsp2.70227
Talk it over
Cement leakage occurred in 215 of 486 vertebrae, but the models predict leakage seen on imaging, not symptoms. How would you weigh a good prediction of leakage against not knowing what it means for patients?
Spine surgery study finds AI can flag cement leakage risk before vertebroplastyhttps://truvace.com/item/development-of-a-multimodal-model-combining-imaging-and-clinical-data-to-predict
Cement leakage occurred in 215 of 486 vertebrae, but the models predict leakage seen on imaging, not symptoms. How would you weigh a good prediction of leakage against not knowing what it means for patients?
Answer in the threadThe external cohort had 50 vertebrae and the prospective cohort 60, with AUCs from 0.722-0.929 in one and 0.834-0.925 in the other. What can cohorts this small tell us about how the models would work in other hospitals?
Answer in the threadThe authors suggest high predicted risk could prompt slower injection or tighter cement volume control. Who should decide whether surgeons change technique, and what evidence should they have first?
Answer in the thread
If you have more time
CT scans came from different scanners and protocols, and removing three scanner-sensitive radiomics features lowered the SVM AUC. How would you check whether a model holds up on a different hospital's scanners?
Answer in the threadIf you were a patient facing vertebroplasty, what would you want to be told about a predicted leakage risk, and what would you want to ask the surgeon?
Answer in the thread
ace
What readers say