Application of Wearable Insole Sensors in In-Place Running: Estimating Lower Limb Load Using Machine Learning

Musculoskeletal injuries induced by high-intensity and repetitive physical activities represent one of the primary health concerns in the fields of public fitness and sports. Musculoskeletal injuries, often resulting from unscientific training practices, are particularly prevalent, with the tibia being especially vulnerable to fatigue-related damage. Current tibial load monitoring methods rely mainly on laboratory equipment and wearable devices, but datasets combining both sources are limited due to experimental complexities and signal synchronization challenges. Moreover, wearable-based algorithms often fail to capture deep signal features, hindering early detection and prevention of tibial fatigue injuries. In this study, we simultaneously collected data from laboratory equipment and wearable insole sensors during in-place running by volunteers, creating a dataset named WearLab-Leg. Based on this dataset, we developed a machine learning model integrating Temporal Convolutional Network (TCN) and Transformer modules to estimate vertical ground reaction force (vGRF) and tibia bone force (TBF) using insole pressure signals. Our model's architecture effectively combines the advantages of local deep feature extraction and global modeling, and further introduces the Weight-MSELoss function to improve peak prediction performance. As a result, the model achieved a normalized root mean square error (NRMSE) of 7.33% for vGRF prediction and 10.64% for TBF prediction. Our dataset and proposed model offer a convenient solution for biomechanical monitoring in athletes and patients, providing reliable data and technical support for early warnings of fatigue-induced injuries.