Hybrid Machine Learning-Neuromusculoskeletal Modeling for Control of Lower Limb Prosthetics

Objective: Current limitations in Electromyography (EMG)-driven Neuromusculoskeletal (NMS) modeling for control of wearable robotics are the requirement of both Motion Capture for both an indoor system and numerous EMG electrodes. These limitations make the technology unsuitable for amputees with only proximal muscles, who need optimal prosthetic device control during everyday activities. Therefore, we developed a novel Machine Learning (ML)driven NMS model able to predict lower limb joint torque only from wearable sensors than can be embedded in a prosthetic device. Methods: After the NMS model calibration of a single healthy subject (OpenSim
^®
software and Calibrated EMGInformed Neuromusculoskeletal Modelling CEINMS Toolbox), an additional ML layer (Gaussian Mixture Regressors) was added to the model to replace the MoCap-derived dependent variables with estimations obtained only from wearable sensors. An on-line open-loop Forward Dynamic (FD) simulation of the knee joint is computed and torque trajectories are compared to experimental ones. Results: Estimations of the novel ML-driven Musculoskeletal model were comparable with experimental knee joint torque during typical locomotion tasks. Accuracy results were comparable to standard EMG-driven MS models and errors are below the threshold of Normalized Root Mean Square Deviation ≤ 0.30 recognized in literature. Conclusions: We developed the first concept of completely wearable and subject-specific EMG-driven NMS model control for lower limb prostheses. The possibility to use this NMS model for FD simulations and the estimation of torque reference control avoids the use of current heuristic and overly complex standard controllers for lower limb prostheses. This research, in fact, represents a key step for the definition of a novel human-machine interface able to create a seamless interconnection between human native control and future wearable robotics.