Article Text

Download PDFPDF
P-037 Using artificial intelligence to predict vertebroplasty outcome
  1. N Carter1,
  2. H Asadi2,
  3. H Kok3,
  4. J Maingard2,
  5. G Anselmetti4,
  6. R Chandra1,
  7. J Hirsch5
  1. 1Interventional Neuroradiology Unit, Monash Imaging, Monash Health, Clayton, Australia
  2. 2Interventional Neuroradiology, Department of Radiology, Austin Hospital, Heidelberg, Australia
  3. 3Interventional Radiology, Northern Hospital, Epping, Australia
  4. 4Interventional Radiology, GVM Care and Research Maria Pia Hospital, Turin, Italy
  5. 5NeuroInterventional Spine Service and Neuroendovascular Program, Massachusetts General Hospital, Boston, MA


Background Vertebral compression fractures (VCFs) are a common complication of osteoporosis and spinal neoplasms. Vertebroplasty presents a treatment option for VCFs that cause pain and disability. Predicting pain relief outcomes based on pre-treatment factors may robustly identify patients likely to benefit from the procedure, and guide clinical decision-making. Machine learning (ML) offers the ability to analyze large volumes of clinical data to formulate accurate outcome predictions.

Purpose The aim of this study was to assess the feasibility of ML techniques to develop a model for accurate prediction of pain relief outcomes after vertebroplasty.

Method This was a retrospective study of a prospectively collected multicenter database. All available demographic and procedural factors were integrated into ML models as predictors, including demographic data, VCF etiology, number and location, procedural details and visual analogue scale (VAS) scores pre- and post-procedure. Successful clinical outcome was defined as a reduction by at least 67% (two-thirds) in VAS pain scores. ML algorithms were trained on these data to develop a model capable of categorizing predictors into good and poor outcomes.

Results We analyzed 5785 patients with 20,463 VCFs treated by vertebroplasty. The majority (75%, n=4,312) were female with a median age of 73.5 (range 19–100) years. The most common VCF etiology was primary osteoporosis (55.2%), followed by secondary osteoporosis (14.7%), metastatic disease (13.2%), primary malignancy (5.5%), trauma (6.4%) and other (5.2%). Most patients had two levels treated (23.9%) followed by one (19.2%) and three levels (17.5%). The majority also had pre-procedure MRI (89.4%). Vertebroplasty was highly effective in pain relief with 97.9% (n=5,662) reporting clinically significant reduction in VAS.

ML analysis based on a naïve Bayes model (figure 1) demonstrated that the number of treated levels, primary malignancy as etiology, pre-procedure brace use, thoracolumbar, cervical levels and metastatic disease predicted a good clinical outcome in decreasing order of confidence. Overall, the ML model was approximately 90% accurate in predicting the outcome for individual patients after VCF.

Conclusion Supervised ML algorithms have promising potential to predict the outcome following vertebroplasty.

Our results suggest that a robust computation model could optimize the decision-making process for VCF management. Incorporating additional data from larger multicenter databases could further improve the accuracy of outcome prediction through iterative learning and refinement of the ML model.

Disclosures N. Carter: None. H. Asadi: None. H. Kok: None. J. Maingard: None. G. Anselmetti: None. R. Chandra: None. J. Hirsch: None.

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.