Physics-Informed RF Features for Bone Fracture Detection and Healing Classification: A Grouped Cross-Validation Study Using 2D-FDTD Simulation Across Five Anatomical Sites

Bone fractures affect approximately 178 million individuals annually worldwide, yet access to radiographic diagnosis remains limited in many clinical and resource-constrained settings. Using simulation-only data, this study investigates radio frequency (RF) sensing as a noninvasive, nonionizing, and portable complement to radiography for fracture screening and longitudinal healing assessment. A comprehensive 2-D finite-difference time-domain (2D-FDTD) electromagnetic framework was developed across three operating frequencies (1.0, 2.4, and 4.0 GHz), five anatomical sites (femur, tibia, ankle, wrist, and clavicle), three fracture-gap widths, and four fracture-healing stages. The term “physics-informed RF features” denotes signal descriptors selected because they correspond to electromagnetic propagation mechanisms in heterogeneous tissue: attenuation and scattering in the time domain, dielectric dispersion in the frequency domain, and fracture-induced disruption of spatial coherence across sensors. A total of 201 such features were extracted from seven-channel RF signals and evaluated using leakage-aware grouped cross-validation at the finite-difference time-domain (FDTD)-configuration level. Systematic evaluation on 15 600 noise-augmented samples showed that random forest achieved the highest binary fracture-discrimination performance (AUC-ROC = 0.832\,± \,0.046 ), whereas random undersampling boosting (RUSBoost) provided the most clinically favorable high-sensitivity operating point (99.0% sensitivity and 45.2% specificity). Leave-one-site-out (LOSO) evaluation demonstrated robust multisite transferability (AUC-ROC = 0.867\,± \,0.006 ), indicating that the learned RF signatures generalize across anatomically diverse skeletal locations without site-specific calibration. For healing assessment, direct five-class staging achieved 37.6%\,± \,2.8% accuracy. A clinically motivated three-class formulation (healthy, active healing, and advanced healing) achieved 61.2%\,± \,2.9% accuracy, corresponding to a 62.8% relative improvement over direct five-class staging, whilst targeted clinical questions achieved AUC-receiver operating characteristic (ROC) values of 0.77–0.83. Because no directly comparable external RF-based multisite grouped-validation healing-stage benchmark is available, state-of-the-art comparison is reported against the closest RF bone-assessment studies. Feature-importance and ablation analyses showed that time-domain statistics (63.5%) and cross-sensor correlations (25.0%) carry the dominant discriminative information, consistent with the physical hypothesis that fractures perturb RF amplitude and spatial coherence. These findings support RF sensing not as a replacement for radiographic imaging, but as a computationally validated candidate for rapid, radiation-free triage and interval monitoring, requiring phantom and clinical validation before deployment.