Abstract
Background/Objectives: Symptomatic intravertebral vacuum cleft (SIVC) is a complication of vertebral compression fractures (VCFs) that leads to persistent pain and deformity. Its prediction remains challenging due to multifactorial causes. Paraspinal muscle fat infiltration has been associated with spinal fracture outcomes but has not been extensively explored in SIVC prediction. Our aim was to develop machine learning (ML) models for predicting SIVC and to evaluate the role of muscle-related variables in improving predictive performance. Methods: Demographic, radiological, and muscle-related variables were collected. ML models-including Logistic Regression, Random Forest, XGBoost, and Multi-Layer Perceptron-were trained and tested under two input conditions: baseline variables (SETTING_1) and baseline plus muscle-related variables (SETTING_2). Model performance was evaluated using accuracy, the area under the receiver operating characteristic curve (AUC), and feature importance analysis. Results: The Random Forest model in SETTING_2, which incorporated muscle-related variables, achieved the highest accuracy (96.6%) and AUC (0.956). Multifidus fatty infiltration (MFfi), erector spinae fatty infiltration (ESfi), and endplate CSA were identified as the most significant predictors. The inclusion of muscle-related variables significantly improved the predictive performance of all ML models. Conclusions: ML models, particularly Random Forest, demonstrated high accuracy in predicting SIVC when muscle-related variables were included. Paraspinal muscle fat infiltration is a critical predictor of SIVC and should be integrated into risk assessment strategies to improve early diagnosis and management.