Special Poster Session 50th International Society for the Study of the Lumbar Spine Annual Meeting 2024

Prediction of Symptomatic Intravertebral Vacuum Cleft Sign with Machine Learning Algorithms - Risk Factor findings and Muscle Variables Strategy (#SP-7c)

Joonghyun Ahn 1 , Young H Kim 2 , Kee Y Ha 3 , June Lee 4 , Youjin Shin 4
  1. Orthopedic Surgery, Bucheon St. Mary's Hospital, College of Medicine, The Catholic University of Korea, Bucheon, Gyunggi-do, Repulibc of Korea
  2. Orthopedic Surgery, Seoul St. Mary's Hospital, College of Medicine, The Catholic university of Korea, Seoul, Republic of Korea
  3. Orthopedic Surgery, Kyung Hee University Hospital at Gangdong, College of Medicine, Kyung Hee University, Seoul, Repulic of Korea
  4. Data Science, The Catholic University of Korea, Bucheon-si, Republic of Korea

Introduction

Vertebral compression fracture (VCF) is a prevalent spinal disorder, which can develop symptomatic intravertebral vacuum cleft (SIVC) representing nonunion of fracture with back pain and postural kyphosis often requiring surgical intervention. Accurately predicting SIVC is crucial to guide clinical decision-making, but this is challenging due to the multifactorial nature of this pathogenesis. Therefore, this study primarily aimed to predict SIVC using a variety of machine learning approaches with clinical data, radiographic compression rates, fracture vertebra kyphotic angle, and multifidus(MF) and erector spinae (ES)-related data obtained from magnetic resonance imaging (MRI). And the second objective of this study is to compare the results predicted only by clinical information and data that can be obtained from plain radiographs with those predicted, including MF and ES related variables, to find out how much muscle (MF and ES) related variables contribute to SIVC prediction.

Methods

Between March 2013 and February 2023, 726 consecutive patients diagnosed with vertebral compression fracture at our institution, who were examined by plain radiographs and magnetic resonance images (MRI), were included in this study. XGBoost, logistic regression (LR), multi-layer perceptron (MLP), and random forest (RF) were selected as machine learning algorithms to predict SIVC. To effectively learn less data, we used SMOTE and performed cross-validation to improve the generalization performance of the model, and the datasets were stratified into training (726, 70%), and testing (218, 30%).

 

Results

Age, sex, adrenal insufficiency, hyper- and hypothyroidism, long-term use of steroid, diabetes, hypertension, endplate cross sectional area, fat infiltration (%) of multifidus (MF) and erector spinae (ES), relative MF and ES, compression ratio, initial angle of compression fracture vertebral body were significantly different in the training and test sets (p < 0.05) between the non-SIVC and SIVC group. In the none muscle variable group, the area under receiver operating characteristics (AUROCs) for XGBoost, LR, MLP, RF were 0.572 (0.517-0.628), 0.663 (0.563-0.763), 0.619 (0.539, 0.699) and 0.617 (0.556, 0.678), respectively, and the accuracies were 0.778, 0.683, 0.700, 0.733. LR model achieved the best AUC performance and XGBoost model achieved the best accuracy performance. In the muscle variables group, the area under receiver operating characteristics (AUROCs) for XGBoost, LR, MLP, RF were 0.833 (0.788-0.878), 0.825 (0.793-0.857), 0.813 (0.763, 0.863) and 0.790 (0.721, 0.860), respectively, and the accuracies were 0.939, 0.892, 0.932, 0.942. XGBoost model achieved the best AUC performance and RF model achieved the best accuracy performance.

Discussion

Including muscle-related variables was successful in prediction and the best performing model was XGBoost. This result has shown that it can be used to prevent SIVC in spinal compression fractures.