Vertebral compression Fracture Risk in Spinal Metastases Patients Following Stereotactic Body Radiotherapy Using Quantitative Imaging Data and Machine Learning

Abstract

Vertebral compression fractures (VCFs) occur in approximately 14% of patients with spinal metastases following treatment with Stereotactic Body Radiotherapy (SBRT). The Spinal Instability Neoplastic Score (SINS) is the current clinical standard for assessing potential mechanical instability in these patients; however, it has several limitations such as it is manually assessed, has an inconsistent relationship with fracture risk and is only semi-quantitative. This study used quantitative CT imaging biomarkers derived from SBRT treatment planning imaging and machine learning models (Logistic Regression, Random Forest, XGBoost, SVM, Gradient Boosting, AdaBoost, Neural Network) to predict vertebral compression fractures (VCF) following SBRT in spinal metastases patients (in 300 thoraco-lumbar vertebrae from 179 patients). The Random Forest model achieved the best performance (sensitivity: 0.64, specificity: 0.76, F1-score: 0.47), showing a 36% improvement in balanced accuracy over SINS. Feature importance analysis identified the quantitative imaging biomarkers, spinal alignment and bone lesion composition (lytic or blastic disease) as the strongest predictors. ML models demonstrated meaningful improvements over existing SINS assessment.