Predicting Fractures in Adjacent Vertebrae Post-Vertebroplasty Using a Machine Learning Approach

Background: Vertebroplasty demonstrates promising clinical outcomes owing to its straightforward surgical procedure, minimal complications, and rapid pain alleviation (1). Nevertheless, post-treatment, 25% of patients encounter vertebral fractures, with 50-67% of such instances observed in neighboring augmented vertebrae (2).
Purpose: This study aims to employ machine learning methodologies and classification techniques to construct predictive models based on identified variables, with the goal of forecasting fractures in adjacent vertebrae subsequent to vertebroplasty.
Methods: This retrospective investigation aimed to identify potential determinants impacting the efficacy of vertebroplasty. Utilizing data from 84 patients diagnosed with osteoporotic vertebral compression fractures who underwent vertebroplasty, predictive models were constructed. K-nearest neighbors (KNN) and logistic regression (LR) algorithms were employed for the prediction of fractures occurring at the adjacent level of the augmented vertebra post-vertebroplasty. The accuracies of these models were documented.
Results: The developed models demonstrated high accuracies, with LR achieving 0.94, and KNN achieving 0.88. Bone mineral density (BMD), cement volume, and cement location emerged as the most significant features in LR. Furthermore, LR exhibited superior performance in terms of macro average (0.92) and weighted average (0.95) accuracy metrics.
Conclusion: The attained high accuracies underscore the efficacy of the developed models in forecasting subsequent adjacent vertebral fractures subsequent to vertebroplasty. Employing these models facilitates accurate fracture prediction, thereby aiding in the prevention of ensuing complications.