Analysis of using EMG and mechanical sensors to enhance intent recognition in powered lower limb prostheses.

OBJECTIVE: The purpose of this study was to determine the contribution of electromyography (EMG) data, in combination with a diverse array of mechanical sensors, to locomotion mode intent recognition in transfemoral amputees using powered prostheses. Additionally, we determined the effect of adding time history information using a dynamic Bayesian network (DBN) for both the mechanical and EMG sensors. APPROACH: EMG signals from the residual limbs of amputees have been proposed to enhance pattern recognition-based intent recognition systems for powered lower limb prostheses, but mechanical sensors on the prosthesis-such as inertial measurement units, position and velocity sensors, and load cells-may be just as useful. EMG and mechanical sensor data were collected from 8 transfemoral amputees using a powered knee/ankle prosthesis over basic locomotion modes such as walking, slopes and stairs. An offline study was conducted to determine the benefit of different sensor sets for predicting intent. MAIN RESULTS: EMG information was not as accurate alone as mechanical sensor information (p < 0.05) for any classification strategy. However, EMG in combination with the mechanical sensor data did significantly reduce intent recognition errors (p < 0.05) both for transitions between locomotion modes and steady-state locomotion. The sensor time history (DBN) classifier significantly reduced error rates compared to a linear discriminant classifier for steady-state steps, without increasing the transitional error, for both EMG and mechanical sensors. Combining EMG and mechanical sensor data with sensor time history reduced the average transitional error from 18.4% to 12.2% and the average steady-state error from 3.8% to 1.0% when classifying level-ground walking, ramps, and stairs in eight transfemoral amputee subjects. SIGNIFICANCE: These results suggest that a neural interface in combination with time history methods for locomotion mode classification can enhance intent recognition performance; this strategy should be considered for future real-time experiments.

Toward the development of a neural interface for lower limb prosthesis control.

Lower limb amputees form a large portion of the amputee population; however, current lower limb prostheses do not meet the needs of patients with high-level amputations who need to perform multi-joint coordinated movements. A critical missing element is an intuitive neural interface from which user intent can be determined. Surface EMG has been used as control source for upper limb prostheses for many years; for lower limb activities, however, the EMG is non-stationary and a new control strategy is required. This paper describes the work completed to date in developing a novel lower limb neural interface.

The use of targeted muscle reinnervation for improved myoelectric prosthesis control in a bilateral shoulder disarticulation amputee.

A novel method for the control of a myoelectric upper limb prosthesis was achieved in a patient with bilateral amputations at the shoulder disarticulation level. Four independently controlled nerve-muscle units were created by surgically anastomosing residual brachial plexus nerves to dissected and divided aspects of the pectoralis major and minor muscles. The musculocutaneous nerve was anastomosed to the upper pectoralis major; the median nerve was transferred to the middle pectoralis major region; the radial nerve was anastomosed to the lower pectoralis major region; and the ulnar nerve was transferred to the pectoralis minor muscle which was moved out to the lateral chest wall. After five months, three nerve-muscle units were successful (the musculocutaneous, median and radial nerves) in that a contraction could be seen, felt and a surface electromyogram (EMG) could be recorded. Sensory reinnervation also occurred on the chest in an area where the subcutaneous fat was removed. The patient was fitted with a new myoelectric prosthesis using the targeted muscle reinnervation. The patient could simultaneously control two degrees-of-freedom with the experimental prosthesis, the elbow and either the terminal device or wrist. Objective testing showed a doubling of blocks moved with a box and blocks test and a 26% increase in speed with a clothes pin moving test. Subjectively the patient clearly preferred the new prosthesis. He reported that it was easier and faster to use, and felt more natural.

The effect of subcutaneous fat on myoelectric signal amplitude and cross-talk.

The effect of subcutaneous fat on myoelectric signal amplitude and cross-talk was studied using finite element (FE) models of electromyogram (EMG) signal propagation. A FE model of the upper arm consisted of skin, fat, muscle and bone tissues in concentric layers. Single muscle fibre action potentials were simulated for muscle fibres at a variety of depths and combined to simulate surface EMG interference patterns. As fat layers of 3, 9 and 18 mm were added to the model, the RMS (root mean square) amplitude of the surface EMG signal directly above the centre of the active muscle decreased by 31.3, 80.2 and 90.0%, respectively. Similarly, surface EMG cross-talk above the region of inactive muscle increased as the fat layer thickness increased. The surface EMG RMS amplitude fell below 5% of its value above the centre of the muscle at 14 degrees, 17 degrees, 34 degrees and 47 degrees from the edge of the active muscle with fat layers of 0, 3, 9 and 18 mm, respectively. An additional model was developed with the subcutaneous fat layer thinned from 9 mm to 3 mm in a small, focal region under a pair of recording electrodes. Reducing the fat layer in this manner caused the surface EMG amplitude at the electrodes to increase by 241% and decreased the EMG cross-talk by 68%; this was near the values for the 3 mm uniform fat layer. This demonstrates that fat reduction surgery can increase surface EMG signal amplitude and signal independence for improved prosthesis control.

Finite element modeling of electromagnetic signal propagation in a phantom arm.

Improving the control of artificial arms remains a considerable challenge. It may be possible to graft remaining peripheral nerves in an amputated limb to spare muscles in or near the residual limb and use these nerve-muscle grafts as additional myoelectric control signals. This would allow simultaneous control of multiple, degrees of freedom (DOF) and could greatly improve the control of artificial limbs. For this technique to be successful, the electromyography (EMG) signals from the nerve-muscle grafts would need to be independent of each other with minimal crosstalk. To study EMG signal propagation and quantify crosstalk, finite element (FE) models were developed in a phantom-arm model. The models were validated with experimental data collected by applying sinusoidal excitations to a phantom-arm model and recording the surface electric potential distribution. There was a very high correlation (r > 0.99) between the FEM data and the experimental data, with the error in signal magnitude generally less than 5%. Simulations were then performed using muscle dielectric properties with static, complex, and full electromagnetic solvers. The results indicate that significant displacement currents can develop (> 50% of total current) and that the fall-off of surface signal power varies with how the signal source is modeled. Index Terms-Control, electromyography (EMG), finite element (FE), modeling, prosthesis.

The hyper-reinnervation of rat skeletal muscle.

This study examines muscle recovery and related changes in the motor unit population of 'hyper-reinnervated' rat skeletal muscle. Medial gastrocnemius (MG) muscles were hyper-reinnervated by either cutting the MG nerve and implanting it on the MG muscle together with additional hind limb nerves, or by crushing the MG nerve and excising the medial portion (50-70%) of the MG muscle. Our findings were that muscles hyper-reinnervated with multiple nerves recovered muscle mass and strength more fully than did the self-reinnervated muscles, more motor units were formed (up to three times the normal number were found), and the mean motor unit size was significantly smaller. A relatively small percentage of muscle fibers became polyneuronally innervated. In contrast, the number of motor units that were formed in the muscle reduction experiments were not significantly larger than was expected considering the mass of the muscles. We conclude that hyper-reinnervation improves muscle recovery, it may be a useful technique for improving function in denervated muscle, and may serve to provide added sources of EMG control signals in some amputees.