Personalized risk assessment of musculoskeletal disorders using artificial intelligence enabled predictive imaging models

Abstract Musculoskeletal disorders (MSDs) impose a substantial burden on mobility, quality of life, and healthcare systems, while early individualized risk stratification remains difficult because structural, clinical, and biomechanical determinants interact nonlinearly. This study presents a reproducible AI-enabled multimodal framework for personalized MSD risk estimation using MRI-like, CT-like, and X-ray-like synthetic imaging surrogates together with synthetic clinical and biomechanical variables. A modality-specific residual CNN architecture extracts imaging representations, structured encoders model non-imaging variables, and a masked gated-attention module performs context-dependent fusion across the five information streams. The computational study uses a fully synthetic cohort of 1,200 subject-level records; no human participants, patient records, hospital-derived images, or human tissue samples are used. Continuous risk estimation and three-class risk stratification are optimized jointly, while Grad-CAM and structured-feature attribution provide model-level interpretability. The reported benchmark results show an accuracy of 92.3%, an F1-score of 0.91, and a cross-validated AUC of 0.94, with an independent-test AUC of 0.93 reported separately. Baseline comparison, ablation analysis, calibration assessment, missing-modality tests, and noise sensitivity analyses support the computational value of gated multimodal fusion. These results provide evidence to show the feasibility of the methodology applied to a controlled synthetic benchmark and create a foundation to be reproduced through other prospective studies with suitable and controlled synthetic cohorts as well as real-world cohorts of acceptable ethical standards.