This study presents the development of an advanced machine learning model based on a two-dimensional (2D) Radio Frequency (RF) sensing framework for refined monitoring of femoral bone fractures. Utilising MATLAB simulations, we created a comprehensive dataset enhanced with variations in bone diameter, muscle thickness, fat thickness, and hematoma size, augmented with multiple sensor configurations (two, four, six, and eight sensors). The model aims to provide a frequent, non-invasive assessment of the fracture healing process compared to conventional imaging methods. Our approach leverages data from six RF sensors, achieving a high overall accuracy of 99.2% in classifying different fracture stages, including "no fracture" and varying degrees of hematoma sizes. The findings indicate that increasing the number of sensors up to six significantly enhances detection accuracy and sensitivity across all fracture stages. However, the marginal improvement from six to eight sensors was not statistically significant, suggesting that a six-sensor configuration offers an optimal balance between performance and system complexity. The results demonstrate significant potential for this technology to revolutionise orthopaedic treatment and recovery management by offering continuous, real-time monitoring without radiation exposure. The proposed system enhances personalised patient care by integrating RF sensing with artificial intelligence, enabling timely interventions and more informed, data-driven treatment strategies. This research lays a robust foundation for future advancements, including three-dimensional modelling and clinical validations, toward the practical implementation of non-invasive fracture monitoring systems.

Accurate angle-of-arrival (AoA) estimation is critical for precise localisation in wearable devices, particularly in challenging wireless environments such as Rician fading with low signal-to-noise ratios (SNRs). This paper proposes a pilot-assisted AoA estimation technique that integrates pseudo-random permutations and Walsh sequences within an OFDM-based transmission framework. The method preserves phase coherence and enhances spatial resolution by optimising pilot allocation and leveraging advanced signal processing. Comprehensive Matlab simulations show sub-degree AoA errors, with robustness validated in ultra-dense multipath scenarios where the Rician K-factor is less than 3 dB. Furthermore, a lightweight, one-dimensional (1D) convolutional neural network (CNN) reduces residual carrier-frequency offsets by over 30%, highlighting a promising synergy between classical signal processing and data-driven learning. Comparative analysis against state-of-the-art techniques and a discussion of computational complexity are provided, underscoring the suitability of the proposed method for next-generation wearable and IoT direction-finding applications.

This paper presents a physics-guided machine learning algorithm for classifying femur fracture presence and subtype using non-ionising radiofrequency (RF) spectral data. Multi-sensor S-parameter responses were generated from a femur phantom model across 1.0–3.0 GHz, producing 104 specimens representing intact bone and three fracture geometries. An exploratory, effect-size-driven band-selection algorithm identified a compact discriminative region between 1.74 and 1.90 GHz. Interpretable classifiers, including k-nearest neighbours (KNN), decision trees, linear discriminant analysis, and Naïve Bayes, were evaluated under strict specimen-level hold-out protocols to prevent data leakage. The KNN algorithm achieved 99.3% frame-level accuracy and 100% specimen-level accuracy for binary fracture detection while maintaining strong robustness in multiclass subtype classification, validated through sensor ablation and leave-one-subtype-out testing. The results demonstrate that compact, interpretable algorithms operating on band-limited RF spectra can achieve reliable, radiation-free fracture classification, supporting future development of continuous and edge-deployable monitoring systems.

Reliable dielectric-property models are a prerequisite for credible forward electromagnetic simulation of microwave-based femoral-fracture monitoring systems. This paper presents a traceable dielectric-property dataset for the 1 to 4 GHz band, with explicit surrogate justification and cross-database validation. Reference targets are consolidated from canonical wideband sources and the IT’IS Tissue Properties Database (v5.0). Compact single-term Cole–Cole parameters are fitted for fracture-relevant tissues and healing-stage constituents. Fracture scenarios (fracture type × healing stage) are constructed using mixing rules with uncertainty bounds derived from physiological variability and volume-fraction uncertainty. We explicitly separate property fitting from forward simulation: the fitted properties are intended as inputs to a forward FDTD solver (with CPML) for design studies; no inversion is performed. Across representative tissues, bandwise agreement between fitted models and reference targets yields NRMSE below 10% with Pearson correlation ρ ≥ 0.97. Cross-database comparison against four independent sources (Gabriel, IT’IS, IFAC-CNR, Sasaki/NICT) demonstrates inter-database consistency within 2.4% maximum deviation at 2 GHz for representative tissues; the additional Cole–Cole fitting approximation introduces deviations of up to 6.9%, which remain within the assigned physiological variability bounds. Surrogate selections for healing-stage constituents are justified physiologically based on water content similarity and assigned conservative uncertainty bounds. As a supplementary downstream demonstration, a machine-learning proof of concept reports (i) 56.1% accuracy on a challenging fine-grained 7-class synthetic task and (ii) 57.1% accuracy on an independent external bio-phantom dataset, confirming that the moderate classification accuracy reflects genuine physical overlap between adjacent healing stages. The dataset and protocols are archived on Zenodo (DOI: https://doi.org/10.5281/zenodo.18027914).