Development of a Multimodal Machine Learning Model for Predicting Second Osteoporotic Vertebral Compression Fractures.
Authors
Affiliations (4)
Affiliations (4)
- Ajou University College of Medicine, Suwon, Korea.
- Korea University College of Medicine, Seoul, Korea.
- Department of Neurosurgery, Ajou University College of Medicine, Suwon, Korea.
- Department of Neurosurgery, Ajou University College of Medicine, Suwon, Korea. [email protected].
Abstract
Osteoporosis increases the probability of osteoporotic vertebral compression fractures (OVCF), and the resultant pain severely limits physical activity and increases mortality and morbidity. This study aimed to develop a high-precision multimodal prediction model to forecast the risk of a second OVCF. This retrospective study included 178 patients from a single institution with a first OVCF between January 1, 2000, and December 31, 2019. The study dataset included 18 preoperative clinical variables, including demographics, medications, comorbidities, bone mineral density (BMD), body mass index, trunk fat/muscle ratio from dual-energy X-ray absorptiometry (DEXA), and imaging data. We employed an intermediate-fusion multimodal approach using deep neural networks, principal component analysis (PCA), and traditional classifiers. The final PCA-fusion+random forest model achieved strong internal cross-validation performance [accuracy 0.979, F1-score 0.969, area under the receiver operating characteristic curve (AUROC) 0.998; mean across folds]. External validation in an independent cohort demonstrated preserved discriminative ability (accuracy 0.904, F1-score 0.902, AUROC 0.923), albeit lower than internal estimates, suggesting dataset shift and possible overfitting. In a sensitivity analysis, a clinical/DEXA-only model also showed good discrimination, but the multimodal fusion model achieved higher recall, F1-score, and AUROC/area under the precision-recall curve, supporting the added value of imaging features. Feature importance and Shapley Additive Explanations highlighted DEXA body-composition measures (e.g., gynoid/android/trunk fat and tissue) and image-derived components as influential. We developed a high-performing multimodal machine learning model for predicting a second OVCF in patients with prior fractures. By integrating clinical and imaging data, the model demonstrated strong predictive accuracy and interpretability, supported by feature importance analysis and visualization.