Towards Remote Gait Analysis: Combining Physics and Probabilistic Models for Estimating Human Joint Mechanics
The connected health movement and remote patient monitoring promise to revolutionize patient care in multiple clinical contexts. In orthopedics, continuous monitoring of human joint and muscle tissue loading in free-living conditions will enable novel insight concerning musculoskeletal disease etiology. These developments are necessary for comprehensive patient characterization, progression monitoring, and personalized therapy. This vision has motivated many recent advances in wearable sensor-based algorithm development that aim to perform biomechanical analyses traditionally restricted to confined laboratory spaces. However, these techniques have not translated to practical deployment for remote monitoring. Several barriers to translation have been identified including complex sensor arrays. Thus, the aim of this work was to lay the foundation for remote gait analysis and techniques for estimating clinically relevant biomechanics with a reduced sensor array.
The first step in this process was to develop an open-source platform that generalized the processing pipeline for automated remote biomechanical analysis. The clinical utility of the platform was demonstrated for monitoring patient gait following knee surgery using continuous recordings of thighworn accelerometer data and rectus femoris electromyograms (EMG) during free-living conditions. Individual walking bouts were identified from which strides were extracted and characterized for patient evaluation. A novel, multifactorial asymmetry index was proposed based on temporal, EMG, and kinematic descriptors of gait that was able to differentiate between patients at different stages of recovery and that was more sensitive to recovery time than were indices of cumulative physical activity.
The remainder of the work focused on algorithms for estimating joint moment and simulating muscle contraction dynamics using a reduced sensor array. A hybrid technique was proposed that combined both physics and probabilistic models in a complementary fashion. Specifically, the notion of a muscle synergy function was introduced that describes the mapping between excitations from a subset of muscles and excitations from other synergistic muscles. A novel model of these synergy functions was developed that enabled estimation of unmeasured muscle excitations using a measured subset. Data from thigh- and shank-worn inertial sensors were used to estimate segment kinematics and muscle-tendon unit (MTU) lengths using physics-based techniques and a model of the musculoskeletal geometry. These estimates of muscle excitation and MTU length were used as inputs for EMG-driven simulation of muscle contraction. Estimates of muscle force, power, and work as well as net joint moment from the proposed hybrid technique were compared to estimates from laboratory-based techniques. This presents the first sensor-only (four EMG and two inertial sensors) simulation of muscle contraction dynamics and joint moment estimation using machine learning only for estimating unmeasured muscle excitations.
This work provides the basis for automated remote biomechanical analysis with reduced sensor arrays; from raw sensor recordings to estimates of muscle moment, force, and power. The proposed hybrid technique requires data from only four EMG and two inertial sensors and work has begun to seamlessly integrate these sensors into a knee brace for monitoring patients following knee surgery. Future work should build on these developments including further validation and design of methods utilizing remotely and longitudinally observed biomechanics for prognosis and optimizing patient-specific interventions.