Lower Limb Locomotor Sensing and Estimation for Athletic Performance and Gait Rehabilitation

Abstract

Mobility is a cornerstone of independence and overall health, yet it is often compromised in both clinical and athletic populations. This dissertation investigates how wearable sensors and data-driven modeling can estimate gait metrics and analyze gait patterns across two distinct use cases: recreational running and post-stroke walking with exoskeleton assistance. In both contexts, the goal is to extend biomechanical assessment beyond the laboratory and support personalized gait monitoring in real-world environments. Inertial measurement units (IMUs) offer a wearable alternative to traditional motion capture systems, enabling kinematic measurements outside of controlled laboratory settings. When paired with machine learning and ground truth force plate data, data from IMUs can estimate kinetic metrics, including joint torques and ground reaction forces (GRFs). Throughout this dissertation, IMUs serve as the primary sensing modality for estimating gait metrics during both running and post-stroke walking across varied environments. We first investigated the feasibility of using data from IMUs to measure overstriding during running, i.e. the horizontal distance between the foot and the body’s center of mass at foot strike. Overstriding is associated with fatigue and running-related injury, and real-world measurement could inform training or technique adjustments. Linear models using sagittal segment angles from 10 recreational runners accurately estimated overstriding during treadmill and overground running and explained over 80% of the variance in peak braking force, demonstrating potential to monitor landing and loading mechanics. Extending this work, we developed machine learning models to estimate kinetic metrics during running. A model trained on treadmill IMU and GRF data from 15 runners estimated braking and propulsion forces during overground running, achieving a root mean squared error (RMSE) of 4.3% bodyweight (%BW), which improved to 2.6%BW with individual fine-tuning using eight strides of overground data. With this method, we demonstrated the transferability of treadmill-trained models to predict braking and propulsion forces during overground running and highlighted the benefit of model individualization. When extrapolated to outdoor track running, the fine-tuned model better captured expected relationships between impulse and speed than the generalized model. These results provide insights into the accuracy and generalizability of kinetic metric estimation from treadmill-based, IMU data-driven models during overground running in and out of the laboratory. Applying this approach in a clinical context, we estimated gait metrics and identified adaptation to exoskeleton assistance post-stroke during human-in-the-loop optimization (HILO) in a community setting. Treadmill-trained models achieved RMSEs of 0.98° for trailing limb angle, 0.076 Nm/kg for peak ankle torque, and 1.2%BW for peak propulsion during overground walking. Results revealed heterogeneous responses to exoskeleton assistance during HILO: some participants increased walking speed with greater paretic limb engagement, while others showed limited gait changes or benefit from assistance. These findings underscore the importance of understanding individual gait adaptation patterns to personalize and optimize rehabilitation. This dissertation concludes with a review of lower limb exoskeletons for locomotor assistance, emphasizing key research priorities for advancing gait rehabilitation. Developing accessible tools for gait analysis and providing personalized feedback or assistance through wearable technologies will be essential for effective gait training and performance evaluation. Looking ahead, the integration of scalable wearable sensing with robust estimation models holds promise for transforming how mobility is monitored and enhanced across clinical, athletic, and everyday mobility use cases.