Running is a fundamental human movement and the analysis of running technique is one of the most widely studied topics in sports biomechanics. Gait analysis labs often pair ground reaction forces (GRFs), measured from force plates, with video-based motion capture to assess the body’s movements (kinematics), as well as the internal and external forces (kinetics) that occur. Despite offering valuable biomechanical insights into running performance and injury risk, these methods are costly and require specialised equipment, limiting both runner accessibility and assessment environments. Wearable sensors offer a low-cost, accessible alternative that can be used to collect data ‘in the wild’, but their accuracy and feasibility need thorough evaluation before large-scale deployment. This PhD thesis evaluates the accuracy of machine learning models in estimating lower limb kinetics during running using a commercially focused pressure insole system paired with additional inertial measurement unit (IMU) sensors at common commercial sensor locations in a large and diverse population. Within the first study, 50 runners completed a 21-minute instrumented treadmill protocol across various speeds and gradients. Data from the NURVV Run system (pressure insoles and a shoe-mounted IMU) served as input for the machine learning models to estimate force plate GRFs, achieving a mean relative Root Mean Squared Error (rRMSEs) of 6.0% (vertical) and 7.1% (anteroposterior). The second study used the same dataset, in which data from the NURVV system and additional IMU sensors positioned at other common commercial wearable locations served as inputs to estimate net joint moment signals. A model with just data from the NURVV system and an IMU located at a typical heart rate monitor location achieved a mean rRMSE of 13.5%, 9.8%, and 7.7% for the estimation of net joint moment signals at the hip, knee, and ankle joints, respectively. The third study evaluated how well the previously created treadmill GRF estimation models would work during overground running, and evaluated techniques that could make use of both the large treadmill dataset and a smaller overground dataset to maximise estimation accuracy. A model trained on both datasets was able to achieve a mean rRMSE of 6.1% and 5.5% for the estimation of overground vertical and anteroposterior GRF signals, respectively. These results show that an affordable, commercially focused wearable sensor system can estimate GRFs and, when combined with data from another common sensor location, net joint moments. Such systems enable large-scale biomechanical data collection beyond lab constraints including outdoor running. They can also support detailed skeletal and muscular loading analysis, or continuous monitoring of lower limb kinetics during training and competition.
Development of wearable technology and innovative methods for the estimation of musculoskeletal loads and the prevention of injury in running: (Alternative Format Thesis)
Carter, J. (Author). 25 Jun 2025
Student thesis: Doctoral Thesis › PhD