Research on Multi-dimensional Feature Fusion Model for Osteoporosis Risk Assessment Based on Deep Learning

Objective Osteoporosis is a progressive metabolic bone disorder characterized by reduced bone mass and microarchitectural deterioration, leading to increased skeletal fragility and susceptibility to fracture. Conventional diagnostic and risk-assessment approaches, such as dual-energy X-ray absorptiometry (DXA) and the FRAX® algorithm, remain limited because they rely primarily on bone mineral density (BMD) and a restricted set of clinical factors, failing to capture the multidimensional determinants of bone strength. This study aimed to develop and validate a deep learning-based multi-dimensional feature fusion model that integrates heterogeneous biological, structural, functional, and genetic information to improve the early identification of osteoporosis and enhance fracture risk prediction. Methods A total of 12 856 participants were aggregated from three major data repositories: the International Osteoporosis Foundation database, a clinical research database on osteoporosis, and a large-scale medical informatics dataset. A unified data-extraction protocol was applied to ensure cross-database harmonization, followed by quality control, variable standardization, and missing-data handling using multiple imputation by chained equations (MICE). A multimodal deep learning framework was constructed to integrate six categories of features: BMD measurements, quantitative bone microarchitecture parameters, bone turnover biomarkers, established clinical risk factors, osteoporosis-related genetic polymorphisms, and sensor-derived balance and gait metrics. A multi-task learning strategy was adopted to simultaneously predict osteoporosis status and 10-year fracture probability. Model training used five-fold cross-validation, and external validation was conducted in an independent clinical cohort. Model performance was benchmarked against DXA alone and the FRAX tool. Results In the internal test cohort, the proposed model achieved an AUC of 0.936 (95% CI: 0.927-0.945), with a sensitivity of 87.5% and a specificity of 91.2%, significantly outperforming DXA alone (AUC=0.889) and FRAX (AUC=0.842) (both PAUC of 0.918 (95% CI: 0.905-0.931) and demonstrated strong calibration (Brier score=0.087). SHAP analyses revealed that, beyond BMD, key predictors included trabecular separation, serum C-terminal telopeptide of type I collagen, balance-related metrics, gait speed, and specific SNPs within the RANKL and VDR loci. A simplified model incorporating only BMD, clinical features, and bone turnover markers preserved high accuracy (AUC=0.917), underscoring its feasibility for resource-limited clinical environments. Conclusion The deep learning-based multi-dimensional feature fusion model markedly enhances the precision and individualization of osteoporosis assessment compared with traditional tools. By integrating biological, structural, metabolic, genetic, and functional dimensions of bone health, the model provides a comprehensive representation of skeletal integrity and robustly improves both diagnostic accuracy and fracture risk prediction. Its strong generalizability across demographic subgroups highlights its clinical applicability. This work offers a promising direction for developing next-generation intelligent decision-support systems that may meaningfully improve osteoporosis screening, risk stratification, and preventive care.