Hybrid Ensemble Deep Learning Framework for Osteoporosis Prediction Using Clinical and CT Imaging Features

Abstract

Osteoporosis, which is a chronic disease of the skeleton characterized by low bone mineral density, micro-architectural deterioration and increased risk of fracture, is asymptomatic until fracture. Therefore, its early prediction is clinically important. A deep-learning template and a CT-image-differentiation framework allow the ensemble training of sub-network architectures on separate data sources. The clinical branch utilizes formal imputation, encoding, and standardisation operators followed by an XGBoost classifier, while the imaging branch employs a pre-trained ResNet50 to extract 2048-dimensional descriptors using a residual convolution, batch-normalization and global-average pooling. The two representations are merged at the feature level and classified by a regularized gradient-boosted tree ensemble whose objective, optimal leaf weights and split-gain criterion are fully derived. The experiments in the study will be based on 1958 records of the Kaggle osteoporosis clinical dataset, and 5614 images of the Bone-Lab CT dataset. The clinical-only model achieves accuracy, precision, recall, F1 and AUC of 0.8724, 0.9620, 0.7755, 0.8588 and 0.8933; the hybrid model achieves 0.8724, 0.9506, 0.7857, 0.8603 and 0.9083. The hybrid model improves AUC by 0.0150 (+1.68% relative), indicating stronger threshold-independent discrimination and improved diagnostic reliability. Index Terms—Osteoporosis prediction, XGBoost, ResNet50, CT imaging, clinical features, ensemble learning, multimodal fusion, deep learning, gradient boosting, medical imaging.