Modeling Ankle Joint Dynamics Based on Inertial Sensor Data Using Machine and Deep Learning

Abstract

In recent years, human motion analysis using inertial measurement units (IMUs) has emerged as a lightweight, low-cost, and portable alternative to advanced biomechanical equipment such as optical motion capture systems or force plates. IMUs enable the recording of motion data in real-world and daily settings, making them powerful tools for medical, sports, and rehabilitation applications. This study proposes a compact and intelligent framework for modeling ankle joint dynamics using only a single IMU mounted on the right ankle, with the dual aim of classifying movement activities and estimating selected dynamic parameters. Data were obtained from the publicly available HuGaDB dataset, comprising accelerometer and gyroscope signals sampled at 100 Hz. The signals were segmented using sliding windows, and 36 statistical features were extracted from each segment. Significant differences between classes were examined using the Kruskal–Wallis test, and the most relevant features were selected with the SelectKBest algorithm. Three modeling approaches were evaluated: Gaussian Process Regression, Random Forest, and a hybrid Convolutional Neural Network with Long Short-Term Memory. The CNN+LSTM model achieved the highest classification accuracy at 96.41%, outperforming the other models, while Random Forest reached 90.5% accuracy using the top 10 selected features. GPR produced satisfactory results for continuous parameter estimation. Training time, resource consumption, and model size were also assessed. The findings demonstrate that high accuracy in human motion analysis can be achieved using data from a single IMU. The proposed framework holds promise for clinical, sports, and rehabilitation contexts, as well as for developing lightweight and personalized wearable systems.