Dynamic Calibration of EMG Signals for Control of a Wearable Elbow Brace

Musculoskeletal injuries can severely inhibit performance of activities of daily living. In order to regain function, rehabilitation is often required. Assistive devices for use in rehabilitation are an avenue explored to increase arm mobility by guiding therapeutic exercises or assisting with motion. Electromyography (EMG), which are the muscle activity signals, may be able to provide an intuitive interface between the patient and the device if appropriate classification models allow smart systems to relate these signals to the desired device motion. Unfortunately, there is a gap in the accuracy of pattern recognition models classifying motion in constrained laboratory environments, and large reductions in accuracy when used for detecting dynamic unconstrained movements. An understanding of combinations of motion factors (limb positions, forces, velocities) in dynamic movements affecting EMG, and ways to use information about these motion factors in control systems is lacking. The objectives of this thesis were to quantify how various motion factors affect arm muscle activations during dynamic motion, and to use these motion factors and EMG signals for detecting interaction forces between the person and the environment during motion. To address these objectives, software was developed and implemented to collect a unique dataset of EMG signals while healthy individuals performed unconstrained arm motions with combinations of arm positions, interaction forces with the environment, velocities, and types of motion. An analysis of the EMG signals and their use in training classification models to predict characteristics (arm positions, force levels, and velocities) of intended motion was completed. The results quantify how EMG features change significantly with variations in arm positions, interaction forces, and motion velocities. The results also show that pattern recognition models, usually used to detect movements, were able to detect intended characteristics of motion based solely on EMG signals, even during complex activities of daily living. Arm position during elbow flexion--extension was predicted with 83.02 % accuracy by a support vector machine model using EMG signal inputs. Prediction of force, the motion characteristic that cannot be measured without impeding motion, was improved from 76.85 % correct to 79.17 % accurate during elbow flexion--extension by providing measurable arm position and velocity information as additional inputs to a linear discriminant analysis model. The accuracy of force prediction was improved by 5.2 % (increased from 59.38 % to 64.58 %) during an activity of daily living when motion speeds were included as an input to a linear discriminant analysis model in addition to EMG signals. Future work should expand on using motion characteristics and EMG signals to identify interactions between a person and the environment, in order to guide high level tuning of control models working towards controlling wearable elbow braces during dynamic movements