Continuous monitoring of human movement is necessary to adapt personalized interventions, evaluate intervention efficacy, and facilitate research in cumulative-load dependent phenomena (e.g., muscle hypertrophy, osteoarthritis). Wearable sensors provide the hardware solution, but a minimal sensor set is required for practical deployment. This presents an analytical hurdle for use of physics-based simulation to calculate the biomechanical variables of interest; a minimal sensor set provides insufficient information. Machine learning techniques have been proposed as a potential solution but at the expense of generalizability and interpretability. Thus, we developed a hybrid approach that utilizes the best of both worlds: machine learning is used only to provide the missing information necessary to drive a physics-based simulation.

We developed an algorithm for simulating muscle contraction during gait using only wearable sensors. To facilitate practical deployment, our method uses a reduced sensor array: two IMUs (one each on the thigh and the shank) and three surface electrodes to measure surface electromyograms of the lateral and medial gastrocnemius and vastus medialis. The musculoskeletal kinematics are computed using the IMU data and optimal state estimation. Machine learning is used only to estimate the excitation of the non-instrumented muscles. Muscle contraction is then simulated using EMG-driven techniques.

Our validation study (https://ieeexplore.ieee.org/document/9507535) demonstrated that our algorithm performed similarly to state-of-art techniques (both physics- and data-driven approaches) in characterizing muscle and joint dynamics in walking gait.

Code and an example dataset is publicly available and maintained at this GitHub repo: https://github.com/gurchiek/nms-dyn