Gait analysis provides critical insights into neuromuscular and postural control, offering valuable biomarkers for detecting and monitoring movement disorders such as Parkinson’s disease (PD). However, conventional laboratory-based systems, which us...
Gait analysis provides critical insights into neuromuscular and postural control, offering valuable biomarkers for detecting and monitoring movement disorders such as Parkinson’s disease (PD). However, conventional laboratory-based systems, which use force plates or motion capture, are costly, non-portable, and unsuitable for continuous real-world monitoring. To address these limitations, this research presents the design, development, of an Internet of Medical Things (IoMT)-based smart insole system for comprehensive gait and Center of Pressure (CoP) analysis. The proposed system integrates eight force-sensitive resistors (FSRs), an inertial measurement unit (IMU), and Bluetooth Low Energy (BLE) connectivity within a cost-effective, portable framework.
Accurate identification of gait phases plays a vital role in Gait analysis. However, most existing smart insole systems depend on high-frequency data acquisition, which increases both power consumption and computational complexity. To address these challenges, this study smart insole platform capable of operating at a low sampling rate of 5 Hz. Each insole enabling synchronized acquisition of plantar pressure and motion data. Fourteen healthy participants completed five walking trials along a 10m walkway, during which synchronized insole sensor data and RGB video recordings were obtained, serving as ground truth for gait-phase annotation. A comparative evaluation of five distinct feature configurations was performed using six conventional Machine Learning classifiers under a participant-wise cross-validation scheme. Among these, pressure-only features analyzed with a Support Vector Machine (SVM) achieved the highest macro F1-score of 0.915, demonstrating that low-frequency plantar pressure signals can effectively differentiate gait phases without significant accuracy degradation. Conversely, models trained solely on IMU-derived features exhibited considerably lower performance, indicating that inertial signals provide limited discriminative information at low sampling rates. Additionally, a visual analytics framework was developed to enhance interpretability, producing spatial plantar pressure heatmaps and activation-frequency maps that distinctly represent load distribution patterns across the four gait phases.
Beyond this work, a multi-domain CoP feature analysis was conducted using the WearGait-PD dataset to distinguish PD participants from healthy controls during the Timed Up and Go (TUG) test and the 3 m walking segment. Data from 39 PD and 38 control participants were analyzed, extracting 144 positional, dynamic, frequency-domain, and stochastic features, including per-foot averages and asymmetry indices. Five classifiers were evaluated using an 80/20 participant-level split with five-fold cross-validation. In the TUG task, Logistic Regression (LR) achieved the best performance (accuracy = 0.875, ROC-AUC = 0.922) with 23 selected features, outperforming the SVM-RBF, Random Forest (RF) k-nearest neighbors (k-NN), and Gaussian naïve Bayes (NB) models.