Int J Inj Contr Saf Promot. 2025 Nov 1:1-24. doi: 10.1080/17457300.2025.2572095. Online ahead of print.
ABSTRACT
Wearable systems for knee pathology detection and prosthetic control remain constrained by diagnostic limitations or rigid actuation. This study introduces an integrated two-phase framework combining non-invasive screening with adaptive prosthetic control. Phase 1 employs novel time-frequency features (Enhanced Mean Absolute Value/Enhanced Wavelength), achieving 94.7% abnormality detection accuracy via Extra Trees classifier, a + 3.16% improvement over conventional features, which is validated through 10-fold cross-validation and rigorous statistical testing (Friedman/Nemenyi, 95% confidence intervals). SHAP analysis yields clinician-interpretable thresholds (e.g. Semitendinosus EMAV > 0.3 mV). Phase 2 utilises multimodal fusion (EMG, FSR, IMU) to achieve 99.2% gait phase accuracy with XGBoost, enabling real-time health-adaptive prosthetic control that dynamically modulates: phase-transition timing (400 ms abnormal vs. 300 ms normal), EMG thresholds (0.15 mV vs. 0.10 mV), and motor gains (2.5× vs. 1.0×) based on pathology status. Validated in a LabVIEW-based control environment across variable terrains and speeds, this end-to-end diagnostics-to-control implementation delivers superior screening accuracy (>4.7% gain vs. deep learning) while enabling context-aware prosthetic adaptation, establishing a new paradigm for accessible musculoskeletal rehabilitation.
PMID:41175030 | DOI:10.1080/17457300.2025.2572095
